Anthropic filed for what could be one of the largest IPOs in history, while Florida decided it’s had enough of OpenAI. Uber discovered that telling engineers to use AI “as much as possible” has consequences in the P&L and Meta has been hacked big time through their chatbot.

Grab your coffee and let’s dive into week 23.

» Stat of the Week: $965 Billion

» Industry Highlights: Market & Product

Anthropic filed a draft registration statement with the SEC on June 1, setting the stage for an IPO that could land as early as October and raise up to $60 billion.¹ Days later, the company tapped Morgan Stanley, Goldman Sachs, and JPMorgan Chase as lead underwriters for an IPO.² At just under a trillion dollars, Anthropic is entering public markets at a valuation that would make most sovereign wealth funds blush. The timing is notable: SpaceX is set to list in June, OpenAI has its own IPO ambitions, and whoever moves first sets the benchmark for everything that follows in the AI sector.

On the infra side, NVIDIA used Computex 2026 in Taipei to announce that its Vera CPU, is in full production, with Anthropic, OpenAI, SpaceX, Oracle, and ByteDance already in the early adopter lineup.³ Jensen Huang also unveiled RTX Spark, a new Arm-based superchip for Windows PCs that combines a Grace CPU and Blackwell GPU in a single package, with Dell, HP, Microsoft, Lenovo, ASUS, and MSI building systems around it.⁴ NVIDIA has spent years owning the GPU market for training; it’s now explicitly targeting inference, agents, and the laptop on your desk. Intel and Qualcomm should be paying close attention.

Brian Chesky, Airbnb’s CEO and co-founder, is reportedly planning to launch his own AI lab while remaining at the helm of Airbnb.⁵ He won’t lead the lab directly, but will back it as a founding supporter.

Meta’s AI chatbot became a security liability this week after it emerged that hackers had successfully hijacked high-profile Instagram accounts simply by asking Meta’s AI support chatbot to change account email addresses, no identity verification required.⁶ Meta says the flaw is fixed, but attacks reportedly continued after the company’s announcement.

Uber has capped employee spending on agentic coding tools at $1,500 per month per tool, after burning through its entire 2026 AI coding budget by April.⁷ Some engineers had been generating monthly bills between $500 and $2,000 in token consumption.

» Industry Highlights: Research & Science

NVIDIA’s RTX Spark deserves a closer look beyond the headline. The chip combines Blackwell GPU and Grace CPU architectures via NVLink-C2C in a laptop-class form factor, targeting up to one petaflop of AI compute with 128GB unified memory.⁴ NVIDIA is betting that personal computing will be redefined around AI inference the same way the smartphone redefined it around always-on connectivity.

» Economic Highlights: The World of AI

Florida became the first US state to sue OpenAI and CEO Sam Altman, in an 83-page complaint alleging the company marketed ChatGPT to children while concealing serious risks, disregarding safety warnings, and prioritising winning the AI arms race over user wellbeing.¹

On safety and trust, President Trump signed a scaled-back executive order on AI on June 2, establishing a voluntary framework in which AI companies can submit frontier models for government review up to 30 days before release.⁹ A broader version had been shelved in May after industry pushback. The final order also directs agencies to create AI cybersecurity benchmarks and an AI cybersecurity clearinghouse. The word “voluntary” is doing a lot of heavy lifting here.

Lastly, GitLab announced it is cutting approximately 350 employees (14% of its workforce) and withdrawing from 22 countries, redirecting savings into what CEO Bill Staples called “GitLab Act 2”: a generational rebuild of its infrastructure to handle 100x growth in agentic AI workloads.⁸ The company reported Q1 FY2027 revenue of $264 million, up 23% year-over-year. It joins a long list of tech companies posting record revenues while simultaneously shrinking headcount, with AI cited as both the reason for the growth and the justification for the cuts.

» Quote of the Week

» Closing

Anthropic is heading to public markets at nearly a trillion dollars, while governments is filing 83-page lawsuits, Meta handed hackers the keys to Instagram, and Uber just put a cap on AI models spending in the company.

Another crazy week for the AI world, so stay tuned for more. The IPO roadshows are being scheduled and they will set the tone for the next year in the technology world.

— Ivo

Footnotes

1 https://techcrunch.com/2026/06/01/florida-sues-openai-sam-altman-in-first-of-its-kind-lawsuit-over-violent-incidents/ https://entrepreneur.com/business-news/anthropic-just-filed-to-go-public-it-could-be-one-of-the-biggest-ipos-ever

2 https://finance.yahoo.com/markets/stocks/articles/anthropic-said-pick-morgan-stanley-175802678.html

3 https://nvidianews.nvidia.com/news/nvidia-unveils-vera-the-cpu-for-agents https://thenextweb.com/news/nvidia-vera-chip-anthropic-openai

4 https://www.nytimes.com/2026/06/01/technology/nvidia-chips-personal-computers.html

5 https://fortune.com/2026/06/04/airbnb-ceo-brian-chesky-plans-to-start-a-new-ai-company/

6 https://thenextweb.com/news/hackers-asked-metas-ai-chatbot-to-hand-over-instagram-accounts-and-it-did

7 https://techcrunch.com/2026/06/02/uber-caps-employee-ai-spending-after-blowing-through-budget-in-four-months/

8 https://techcrunch.com/2026/06/03/gitlab-cuts-14-of-staff-as-it-scales-its-platform-to-serve-ai-workloads/

9 https://www.politico.com/news/2026/06/02/trump-signs-downsized-ai-order-00946389

I have a bad habit of building things that give me unwelcome news about myself. Stay tuned, as that’s what the new Fernão module will be.

If you’re new here: I’ve been building Fernão, my own personal AI assistant. A very sparse stitched together app with different tabs, each one handling a slice of my life. Part 1 was a schedule organizer. Part 2, a task breaker. Part 3, a writing editor (which is, somewhat ironically, helping me write this very post).

And now Part 4: a portfolio analyzer that looked me in the eyes and told me I was doing it wrong.

Time to meet my nemesis — meaning, me as a stock picker.

First, Some Context On Why I Even Needed This

I’ve been investing for 5 years. Nothing crazy as I’m not day trading crypto at 3am or anything. I hold a mix of ETFs, individual stocks, some bonds, a couple of REITs.

At the same time, I like to read about companies, I follow earnings and I have opinions on things like competitive moats and management quality and total addressable markets.

I was, in other words, doing what most retail investors think they can do: putting in a reasonable amount of effort and assuming it was worth something — even if it was just for learning experience.

(And yes, I’ve read all the warnings. “Most retail investors underperform the market.” Moreover, I have a statistics masters. I knew the odds. I just needed to prove my own stupidity through experience. 😬)

When you start to invest, it’s difficult to have a free app that can tell you exactly your cost basis and performance. You can look at your portfolio on a random Tuesday, see that it’s up overall, and conclude that things are going fine, but there aren’t many tools that give you a holistic view of how you are performing through time and against many benchmarks (Yahoo finance gives a somewhat overview of that, but still limited for investors in Euro).

I wanted to check my performance over time and figure out what to do differently. So, having Fernão right here in my computer, I asked for help.

Spoiler alert: my investment decisions are not yielding extraordinary results.

Feeding Fernão: From Spreadsheets to Something Useful

The data source is an Excel file where I log every trade I do with date, ticker, buy or sell, quantity, price. I’ve been logging it since I started investing, which is the one thing I did right from the beginning.

Consistency in data entry is unsexy but it’s what makes a tool like this possible. If you’re not tracking your trades, the first step isn’t building the analyzer, it’s spending a painful afternoon reconstructing your history from brokerage statements (not fun!)

From that file, Fernão builds a few different views of what I own.

(I really need to improve the UI on this one. The emojification of front-ends by Claude is giving me the itch)

(Note: the values, except percentages, that I’ll show in this post are fictional or hidden, for privacy reasons).

The sector breakdown is the first reality check. It shows me where my capital lives. I had sectors I considered “small positions” that had grown, and sectors I thought were core holdings that had been diluted down as I bought other stuff. Seeing it as a pie chart is not the most data-viz recommended way, but they get the job done.

In the screenshot above, we can see that I have 105 positions open (some of these positions are repeated across brokers as I try to diversify across 3 different banks here in Portugal. Also, I have other assets such as real estate or Certificates of Deposit — the real unique positions should be around 50 to 60).

Also, we see the division by three groups:

• Sector: It’s not industry of the stock, but tells me the portion of the economy I’m investing in (World Stocks, US Stocks, Gold, etc.). In hindsight should have called this Asset Type.

• Type: If it’s an ETF, Individual Stocks, Real Estate, etc.

• Market: If it’s negotiated in the Nasdaq, Amsterdam, Deutsche Borse, etc.

Nevertheless, I immediately see that I have too many scattered positions. One of the things that I’ll be aiming to do in the next few months is to simplify my portfolio a lot, and probably reduce the positions to around 25/30 (taking into account duplication). As we will see next, I am suffering from diworsefication (where you diversify too much that your portfolio ends up returning less than the overall market).

This overview section is what it is — context, structure, a picture of the thing. Looking at it, we already see that there are too many positions open and that I should try to simplify in the future.

But, the interesting part is what comes next.

The Analysis Tab

The “Analysis” tab is where Fernão really showed me the important details.

The first thing I asked for was a historical comparison between my portfolio and the S&P 500 , on a Euro basis. This part matters a lot, because I’m based in Portugal. Most investors benchmark themselves against the dollar-denominated S&P 500 when they don’t actually live in dollars.

The only benchmark that’s real for me is the one in the currency I actually spend (Euro), so let’s analyze it:

This graph was the slap in the face.

Up until 2024, I was keeping pace, not beating the index, but close enough. Then the gap opened. My best theory: the stocks I like (blue chips, mature businesses, the boring stuff) aren’t where the market is putting its money right now. AI has a type, and it’s not them. I think it may revert, and I still believe in some of the stocks I bought.

But what bothers me more than the recent gap is the years before it — was where I was on par with the index. From 2022 to 2024, I was spending a lot of time choosing stocks. Hours of research, company analysis, reading filings. And the outcome was: about the same as doing nothing.

This isn’t a new story. Warren Buffett literally bet a million dollars on it — a public wager that a simple S&P 500 index fund would outperform a handpicked basket of hedge funds over ten years. The hedge funds lost. If the smartest, most resourced active managers in the world can’t consistently beat the market, what exactly was I doing thinking I could from my laptop on a Sunday afternoon?

That said — I’d do it again. Not because it was profitable, but because I needed to find out for myself. Some lessons only land when they’re yours! This one’s mine now, properly. There’s so much growth in making mistakes and learning quickly!

There’s this concept in investing called alpha, which is the excess return you generate compared to the benchmark. Positive alpha means you’re adding value with your stock selection. Negative alpha means you’d have been better off buying an ETF and going to the beach.

My alpha was negative.

One year of outperforming the index, the year the index was negative. My selection seems good when the index turns negative, but that doesn’t compensate the very positive years of the index.

Every other year, I lagged. 2023, 2024, 2026 especially, and my average outperformance is around -4.8 p.p. Ouch!

Every year you underperform isn’t just a small miss in that year — it’s a smaller starting point for compounding in the next year, and the one after that, and every year for the next two decades. A consistent 5% annual underperformance is not boring and manageable. Over 25 years it’s the difference retiring in your 40’s or 60’s.

The saving grace is time. I still have a long runway (hopefully) ahead of me, and catching these mistakes means I can do something about it. The portfolio is correctable, the habits are correctable, and compounding still has decades to work in my favour if I stop getting in its way.

Main conclusion from Fernão’s analytics: I was putting in more work than an index investor — more research, more monitoring, more mental energy spent thinking about positions. I was taking on more risk, because individual stocks are meaningfully more volatile than a broad index fund. And in return for all of that extra work and extra risk, I was producing worse outcomes.

Thanks Fernão for pointing out my bad strategy!

The Wins, the Blunders

Diving deep from the macro levels, I also want Fernão to give me an analysis on individual positions.

My top stocks by percentual return are Apple, Digital Ocean and Lam Research. The issue here is that I’ve also had my share of blunders as well:

In this individual analysis, I can now keep track of all my individual positions in the portfolio.

At 60–70 positions, I had created a portfolio that was hard to manage and that was obscuring underperformance by spreading it thin enough to not immediately notice.

So: too many positions, negative alpha, underperforming the index consistently. The diagnosis was pretty clear. Let’s see what my personal assistant recommends next.

Fernão as My Financial Advisor: What Should I Actually Do?

Data without prescription is just expensive introspection.

With the diagnosis of the ill issues around my portfolio, I built a “Consolidate & Invest” tab to turn the analysis into action. The idea is simple: every month, I give Fernão my current portfolio state, my investment goals, my tax situation, and how much new capital I want to deploy, and I ask it what to do.

The prompt took a while to get right. In Generative AI, the quality of the output is directly proportional to the quality of the context you give it, which is true of every AI application, but especially obvious when the subject is your specific financial life. Once I gave it real numbers such as actual allocations, targets, emergency fund balance, monthly expenses, the recommendations got useful.

So, given all of this: Let’s imagine I want to invest €2,000 next month. What does Fernão say?

The executive summary flagged three things. One: my core growth ETFs are significantly underweight. I’d been trickling money into individual stocks and letting the ETF allocation drift down over the years. Two: several individual stocks and some bond and REIT ETFs are both underperforming and sitting above their target allocation, as we’ve seen.

And three, as I had a large expense in my life in the last month, my emergency fund covers 2.8 months of expenses, while the target is 6.

Here is the prompt I’m passing to Fernão for the recommendations:

You are a portfolio advisor helping Ivo Bernardo optimize and grow his portfolio. You have two jobs today:
(1) identify what to clean up / rebalance, and
(2) produce a concrete investment routing plan for this month.

{portfolio_summary}

- -
## IVO’S FULL FINANCIAL PICTURE
### Target Portfolio Allocations (from investing ratios)
| Asset Class | Implied Dividend | Implied Return | Target % | Current % |
| - - - - - - - - - - | - - - - - - - - -| - - - - - - - - | - - - - - | - - - - - -|
| VWCE | 0% | 13% | 35% | 16.48% |
| Nasdaq | 0% | 18% | 25% | 9.75% |
| SPYD | 1.8% | 5.5% | 10% | 9.08% |
| Asia Pacific Div | 2.52% | 2% | 3% | 2.85% |
| Euro Div | 3.24% | 0% | 1% | 1.12% |
| TDIV | 2.52% | 7% | 9% | 0% |
| Individual Stocks | 3% | 6% | 17% | 17% |


## RETURN DATA - CRITICAL INSTRUCTION
The context above includes a section **”ACTUAL POSITION RETURNS (from verified price history)”**. This table contains real YTD / 6M / 1Y / 5Y returns for each position, resolved via the correct ticker mapping (e.g. “Amundi Nasdaq” maps to 6AQQ.DE, not QQQ or ^NDX). **Always use these figures when discussing position performance.** Do NOT substitute generic index benchmarks or make up returns for positions that appear in this table.
 - -
## OTHER CRITICAL CONTEXT
- **High Growth individual stocks** are intentionally small positions for learning/practice. Do NOT recommend selling them unless the thesis is fundamentally broken.
- **Same instrument across multiple banks = intentional custodian diversification.** Never flag this as redundancy.
- **Fragmented ETF positions across banks = intentional.** Only suggest consolidating same-bank duplicates.
- Ivo holds collectibles and real estate as non-correlated alternatives - some equity concentration is acceptable.
- **TDIV** is at 0% vs 9% target - highest priority to start.
- **VWCE** at 16.5% vs 35% target - major underweight, needs the most new capital.
- **Nasdaq** at 9.75% vs 25% target - also significantly underweight.
- **Individual Stocks** is overweight at the overall wealth level (26.6% vs 17% target) but at target within the investing ratios framework. Trim only if genuine redundancies exist.

 - -
## EMERGENCY FUND RULE
The context above includes an **”EMERGENCY FUND STATUS”** section with live data (avg monthly expenses, 6-month target, current DP’s balance, months covered).
- If **⚠️ UNDERFUNDED**: include a 🛡️ EMERGENCY FUND action at the TOP of Part 2, and add a DP’s top-up row to the Part 3 routing table with the shortfall amount. Investment into growth assets is secondary until the fund is secure.
- If **✅ FULLY FUNDED**: briefly acknowledge it in Part 1 and proceed with the normal investment plan.
- Always state the exact figures: “€X/month avg expenses → 6-month target €Y → current DP’s €Z → X.X months covered.”
 - -
## YOUR OUTPUT (4 parts)
### PART 1 - EXECUTIVE SUMMARY
3 sentences: current state vs targets, biggest gaps, what this month’s plan achieves. Include one sentence on emergency fund status.
 - -
### PART 2 - SELL / REDUCE / CONSOLIDATE
If emergency fund is underfunded, start with:
 - -
#### 🛡️ TOP UP EMERGENCY FUND: DP’s / Term Deposits
**Action:** Deposit €[shortfall] into DP’s / term deposits to reach 6-month target
**Shortfall:** €[amount]
**Reasoning:** Current DP’s covers only X.X months of expenses vs the 6-month target. Reinforce before deploying capital to growth assets.
 - -
Then for each position to act on:
 - -
#### 🔴 SELL / 🟡 REDUCE / 🔀 CONSOLIDATE: [Ticker(s)]
**Action:** [One-line action]
**Estimated proceeds:** €[amount]
**Reasoning:** [2–4 sentences. Reference P&L, position size, target gap.]
**Contrarian View:** [1–2 sentences - why someone might keep it]
**Learning Position?:** Yes / No
 - -
Aim for 2–4 actions. Sum up total estimated sell proceeds at the end of Part 2.
 - -
### PART 3 - THIS MONTH’S INVESTMENT ROUTING PLAN
**Total capital to deploy = fresh capital ({invest_amount}) + proceeds from Part 2 sells**
Allocate across ETFs both are underweight and serve the passive income goal.
If emergency fund is underfunded, add a DP’s top-up row at the top:
| Destination | Rationale | € to deploy | Action |
| - - - - - - -| - - - - - -| - - - - - - -| - - - - |
| DP’s (if underfunded) | Emergency fund shortfall | €… | Transfer to term deposit |
| VWCE | Largest ETF underweight | €… | Buy X units on [broker] |
| TDIV | 0% → 9% target, priority | €… | Start position |
| Nasdaq (CNDX/EQQQ) | 9.75% → 25% target | €… | Add to existing |
| Other | (if applicable) | €… | … |
| **TOTAL** | | **€…** | |
End with: **”After this month: estimated new % for VWCE ~X%, TDIV ~X%”**
Be decisive, specific, and data-driven. Use the actual portfolio data above for current values.

And now.. let’s see the recommendations. The financial data has been anonymized, but the type of recommendations Fernão provides should still come through clearly.

PART 1 — EXECUTIVE SUMMARY

Ivo Bernardo’s portfolio, valued at x across x positions, demonstrates strong diversification across asset types and platforms, with a commendable focus on learning through individual stock picks. However, key growth ETFs like VWCE and Nasdaq are significantly underweight, while several individual stocks and bond/REIT ETFs show persistent underperformance and are over their target allocation. This month’s plan prioritizes shoring up the emergency fund, then streamlines underperforming assets to redirect capital towards core growth and passive income generators, accelerating the path to the €x/month passive income goal. The emergency fund is currently underfunded, covering only x months of expenses.

PART 2 — SELL / REDUCE / CONSOLIDATE

🛡️ TOP UP EMERGENCY FUND: DP’s / Term Deposits Action: Deposit x into DP’s / term deposits to reach 6-month target Shortfall: x Reasoning: Current DP’s covers only x months of expenses (€x) versus the 6-month target of €x. Reinforcing this essential safety net is the top priority before deploying capital to growth assets.

🔴 SELL: Sweetgreen (SG) Action: Sell entire position in Sweetgreen (SG). Estimated proceeds: €x Reasoning: This “Stock High Risk, Fast-Growth” position has a current value of only €x, an abysmal P&L of -78.1%, and a 5-year return of -74.2%. Analyst consensus implies a further downside of -21.6%. While intended for learning, the fundamental thesis appears severely broken, making it an inefficient use of capital.

🔴 SELL: RH (RH) Action: Sell entire position in RH (RH). Estimated proceeds: €x Reasoning: This “Fast-Growth” individual stock has experienced a significant P&L loss of -65.6% and a 5-year return of -78.0%. While analysts see a modest implied upside of +6.0%, the persistent underperformance indicates a broken thesis for a learning position.

🔴 SELL: InMode (INMD) Action: Sell entire position in InMode (INMD). Estimated proceeds: €x Reasoning: As a “Stock High Risk, Fast-Growth” position, InMode exhibits a substantial P&L loss of -65.8% and a 5-year return of -70.5%. The position is very small, and its consistent underperformance suggests the learning value is diminished given the capital could be better utilized elsewhere.

🔴 SELL: Ishares Euro Corporate Large Bond (IBCX.DE) Action: Sell entire position in Ishares Euro Corporate Large Bond (IBCX.DE). Estimated proceeds: €x Reasoning: This Bond ETF has a significant P&L loss of -56.3% and a negative 5-year return of -0.9%. The Bond ETF asset class is currently overweight (3.71% vs 2% target), making this an opportune time to reduce exposure to an underperforming asset within this category.

SELL: iShares European Property Yield UCITS ETF (IPRP.AS) Action: Sell entire position in iShares European Property Yield UCITS ETF (IPRP.AS). Estimated proceeds: €x Reasoning: This REIT ETF is significantly overweight relative to its target (0.17% vs 0.01%) and has a P&L loss of -28.0% with a negative 5-year return of -14.9%. Reducing this position aligns with target allocations and removes a consistently underperforming asset.

Total Estimated Sell Proceeds: €x + €x + €x + €x + €x = €x

PART 3 — THIS MONTH’S INVESTMENT ROUTING PLAN

Total capital to deploy = fresh capital (€x) + proceeds from Part 2 sells (€x) = €x. Given the emergency fund’s underfunded status, all available capital this month will be directed to reinforcing it.

After this month: The emergency fund will increase to €x (current) + €x = €x. This covers x months of expenses, significantly closer to the x-month target of €x, leaving a remaining shortfall of €x. Further investment into growth assets will be possible once the emergency fund is fully secured.

Fernão’s recommendations are only as good as the context I give it, and I validate everything before acting. The goal isn’t to outsource the decisio, it’s to make sure I’m not missing something obvious before I make it. But this helps me deal with decision fatigue as I come into conclusion that that decision fatigue was only bringing me underperformance.

The Dividend Tab: Discovering That Income Isn’t Free

Two last tabs before we finish this blog post.

The dividend tab does two things. It projects my forward dividend income based on current holdings and historical payouts. And it applies tax rates, both the withholding at source in each holding’s country and the income tax I’d owe in Portugal on top.

Here’s roughly what happens when I earn dividends in Portugal. A US company pays me €100. The US withholds 15% at source, treaty rate for Portuguese residents. I’m left with €85. Portugal then taxes that €85 as regular income. Depending on the bracket, I net somewhere between €55–€70 of the original €100 after the US withholding credit, I always model dividends at worst-case to keep my projections conservative.

Meanwhile, if I hold an accumulating ETF, one that reinvests dividends rather than distributing them, those same gains compound inside the fund. I only pay tax when I eventually sell, and potentially at a more favourable rate.

The asymmetry is significant over a 20–25 year horizon. I was holding several distributing ETFs when accumulating equivalents existed, earning dividends that were immediately being taxed at a high combined rate, and thinking of it as “income.” I’ll still want to target around 1% of dividend income as time goes by, but I’ll not overweight distributing dividends.

A stock paying 4% today that’s growing its dividend at 10% annually will, after enough years, be yielding a much higher amount on your original cost. I also map this for the future in the application. Here are some examples based on my portfolio positions and how the net dividend projection may become:

Correlation Analysis: What Diversification Looks Like

The correlations tab was the most technically interesting thing to build, and also the most educational to actually use.

The basic idea of diversification is that you want your assets to behave differently from each other. If everything falls together when the market drops, you weren’t diversified, you just had the same bet in different packaging. The goal is to own things whose prices move with some independence, so losses in one part of the portfolio don’t necessarily show up everywhere else at the same time.

Correlation is how you measure this. A correlation of +1 means two assets move in perfect lockstep , if one goes up 10%, the other goes up as well. Zero means no relationship. Negative means they tend to move in opposite directions. In practice, most financial assets have correlations somewhere between 0 and +1. Fernão also gives me this data with a rolling correlation through time — take the example of Google and Apple:

Google and Apple: strongly correlated, as expected. Both are massive US tech companies, both exposed to the same macro environment (interest rates, ad markets, consumer sentiment, regulatory pressure) and both owned heavily by the same institutional investors. When one moves, the other tends to move in the same direction. Holding both gives you company-specific exposure, but it doesn’t help you from a diversification standpoint.

Apple against the total market is where it gets interesting. Apple has become so dominant in major indices, at various points the largest company in the world by market cap, representing a significant chunk of the S&P 500 — that it and the index have become nearly inseparable. The correlation between Apple’s stock and the Total Market is even larger:

Gold is where the picture changes. Against the total stock market, gold’s rolling correlation hovers around zero and regularly swings negative. This is what “flight to safety” looks like:

This is why portfolio managers allocate to gold even though it doesn’t pay a dividend, doesn’t have earnings, doesn’t do anything except sit there and be gold. You’re not holding it for returns, you’re holding it to reduce how badly things go when stocks go down. The drag in good times is the price of the cushion in bad ones.

Bitcoin is the most interesting case study here, and it’s one where the data tells a pretty different story from the marketing. Some pitches for Bitcoin is “digital gold” , scarce, decentralised, a hedge against inflation and currency devaluation. Under that narrative, you’d expect it to behave like gold: low or negative correlation with equity markets, a place capital moves toward when risk appetite contracts.

The rolling correlation chart between Bitcoin and the NASDAQ is… not that. They’re positively correlated. When tech stocks fall, Bitcoin tends to fall too. When tech stocks rally, Bitcoin often rallies alongside them. The likely explanation is that Bitcoin’s investor base and the NASDAQ’s investor base overlap heavily.

For Bitcoin to earn the uncorrelated status that gold actually holds, it would need to decouple from risk sentiment in a sustained, consistent way across multiple market cycles.

Okay But What Am I Actually Changing?

Building this module gave me a lot of insights. I was feeling frustrated as an individual investor of not having a tool to deeply analyze my portfolio — most tools are too expensive or thought for institutional investors. Luckily, the AI world is giving us these abilities to build our own tools quite fast. This is not, of course, an enterprise grade application, but more of a personal cockpit that I’m building.

So let me just say what I’m doing differently after analyzing this data:

• Stopping new capital going into individual stocks. Not selling everything immediately, some positions still have real theses and some would trigger unnecessary tax events if sold now.

• Switching to accumulating ETFs wherever possible. For every ETF where an accumulating version exists, I want the accumulating one.

• Fixing the emergency fund before anything else.

• Reducing position count over time. I have too many open positions. I’m not going to make 90 sell decisions at once, but as positions naturally run off or hit sell criteria.

What I Actually Learned From Building This

I built this module in my Personal Assistant to get better financial information.

I’m lucky I built this when I still have a lot of years of investing ahead of me. The earlier you catch a bad strategy, the more time you have to compound the correct one instead. That’s the whole game, really.

What are your thoughts on using AI for this kind of personal financial analysis? And if you were building your own personal assistant, what module would you build first?

More modules to come!

— Ivo

This was the week AI and theology briefly collided, as Pope Leo had a word for the technocrats.

But the ones truly in need of divine intervention? The startup that (allegedly) accidentally got a bill $500 million after letting employees run Claude Code with zero guardrails. Meanwhile, two CEOs quietly walked back their robot apocalypse prophecies. Funny how IPO season inspires sudden optimism.

Grab your coffee and let’s dive into week 22.

» Stat of the Week: $500 Million

» Industry Highlights: Market & Product

Anthropic launched Claude Opus 4.8¹ this week, an upgrade to its flagship model that brings improvements across coding, agentic tasks, and reasoning benchmarks. Fast mode now runs at 2.5× the speed and costs three times less than previous models.

The launch happened when the company just closed its Series H round, becoming officially the most valued startup in the world.

Tied with the Claude models, an unnamed enterprise client made headlines for all the wrong reasons this week: an AI consultant revealed to Axios that one of their clients burned $500 million in a single month on Claude, simply because nobody thought to cap employee usage². Engineers were reportedly checking the weather with agentic workflows, at hundreds of dollars a pop. The same week, Microsoft confirmed it has been quietly scaling back internal Claude Code licences³, and Uber admitted it blew through its entire 2026 AI budget by April⁴.

Asana acquired Y Combinator-backed Stack AI for a reported $75 million, adding no-code AI agent building capabilities to its project management platform. Asana’s Q1 revenue hit $205 million, yet its stock sits 53% below its January peak, as the company has been a victim of the famous saaspocalypse.⁵

Lastly, DuckDuckGo saw nearly 28% more visits in the week following Google’s announcement that users love AI Mode. Will Google learn that people just want a choice? Or will they prove that when you embed AI into our product successfully, no one cares?⁶

» Economic Highlights: The World of AI

Pope Leo XIV published his first encyclical this week, Magnifica Humanitas, and it uses AI as a lens to diagnose something older and more structural: concentrated power, eroding democracy, and a tech elite that increasingly shapes the world to its own advantage.

The encyclical isn’t really about AI. It’s about who controls the future, and whether anyone is accountable for that. The Economist noted that the document explicitly attacks what it calls “technological messianism”, the quasi-religious belief that enough compute will solve enough problems. Is the pope really concerned about the future of society? Or just afraid to have another competitor in the religious field? ⁷

Speaking about heads of the technological church, Sam Altman and Dario Amodei both spent the week walking back earlier statements about AI causing a jobs apocalypse, a reversal that Fortune connected to the fact that both companies are preparing for blockbuster IPOs. Altman told an Asia-Pacific audience that mass job displacement is unlikely, even as he acknowledged entire categories (customer service, for instance) are disappearing. ⁸ ⁹

» Track of the Week: Greg Brockman in the Knowledge Project Podcast

This week had a little of everything: a pope diagnosing Big Tech's original sin, a company lighting half a billion dollars on fire because someone forgot to add a spending cap, and two AI CEOs discovering that "mass unemployment is coming" is a harder pitch when you're trying to get retail investors to buy your shares.

See you next week!

— Ivo

Footnotes

1 https://www.anthropic.com/news/claude-opus-4-8

2 https://www.msn.com/en-us/news/technology/company-accidentally-blows-500-000-000-on-claude-ai-in-one-month/ar-AA24nahv

3 https://www.theverge.com/tech/930447/microsoft-claude-code-discontinued-notepad

4 https://fortune.com/2026/05/26/uber-coo-ai-spending-tokens-claude-code/

5 https://thenextweb.com/news/asana-acquires-stack-ai-agent-builder-saas

6 https://www.pcgamer.com/hardware/duckduckgos-ai-free-search-saw-nearly-28-percent-more-visits-in-the-week-following-googles-insistence-that-people-love-ai-mode/

7 https://techcrunch.com/2026/05/25/the-popes-ai-encyclical-isnt-really-about-ai/

8 https://thenextweb.com/news/sam-altman-says-an-ai-jobs-apocalypse-is-unlikely

9 https://fortune.com/2026/05/26/sam-altman-dario-amodei-walking-back-ai-jobs-apocalypse-prophecies-ipo/

Some of you may know, but a couple of months ago I wrote a book. It’s called The AI Handbook for the Curious Mind, and it’s available on Amazon or Gumroad. As this newsletter has been gaining some traction lately, I’ve decided to do something a little different for paid subscribers here on The Data Journey.

Every two weeks, I’ll be releasing a full chapter here, exclusively for paid subscribers. I keep looking for ways to give back to the people who support this newsletter, and sharing the book felt like one of the more meaningful ones. 😄

If you want to check other chapters, check:

• Chapter 1: Introduction

• Chapter 2: Artificial Intelligence 101

• Chapter 3.1: Statistics Intro

• Chapter 3.2: Data Types

• Chapter 4: Intro to ML

If you’ve been following this series, you know what’s been building up. We’ve survived statistics. We’ve built a robot that sorts fruit. We’ve drawn the distinction between supervised, unsupervised, and reinforcement learning.

Now we’ll dive into machine learning, starting with one of its foundational algorithms — linear regression. By the end of this section, you’ll understand how models learn from data and use it to make decisions.

Today, we meet our first real machine learning model.

Read more

Good morning everyone,

Week 21 had Google on the spotlight with I/O 2026. We have our bunch of new models and a long awaited change to Google Search.

In the law department, Musk lost his lawsuit, and OpenAI immediately started sprinting toward an IPO.

And on the financial side, Anthropic mentioned it’s about to turn a profit, and NVIDIA posted numbers that could be treated as a typo 10 years ago.

Grab your coffee!

» Stat of the Week: 200 billion

Nvidia CEO Jensen Huang declared this week that AI agents represent a brand new $200 billion market opportunity, specifically for CPUs designed to run them.

» Industry Highlights: Market & Product

A day after Elon Musk lost his lawsuit threatening OpenAI’s structure1, leadership, and finances, the company is reportedly ready to move forward with an IPO2, potentially as soon as September, with Goldman Sachs and Morgan Stanley on the deal and confidential paperwork possibly filed within weeks. The two moves are deeply related, as the lawsuits’ central argument is that OpenAI should have remained a non-profit company (and therefore could not become a public company).

On the NVIDIA side, the company presented its latest financial resuls, with $81.6 billion in revenue, up 20% quarter-over-quarter, ($75.2 billion of it from data centers)3. Also interesting is to see that NVIDIA’s stakes in privately held companies nearly doubled in three months, from $22 billion to $43 billion. The chip vendor is increasingly also the venture investor in the people buying the chips. On the same call, Huang declared that Nvidia’s new Vera CPU opens “a brand new $200 billion TAM,” a market the company has never addressed before.4

Continuing on the US front, Anthropic bought Stainless, the SDK and MCP-server tooling company that has powered every official Anthropic SDK since the earliest days of the API5. The acquisition market continues to heat up.

On the acquisition front, Europe’s flagship lab (Mistral) acquired Vienna’s Emmi AI, a physics-simulation startup, for an undisclosed sum (reportedly near €300M), its second acquisition of 2026 after Koyeb. Emmi builds “large engineering models” that simulate fluid dynamics and structural deformation in real time versus the multi-day runtimes of traditional solvers.6

Finally, let’s look into Google I/O 20267. The company launched Gemini Omni, a “world model” that generates and edits high-fidelity video from any mix of text, audio, and video, and Gemini 3.5 Flash, the first in its newest family combining “frontier intelligence with action.”

Watch the full recap of the event below:

» Industry Highlights: Research & Science

An AI cracked an 80-year-old Erdős problem: OpenAI model disproved a central conjecture in discrete geometry8, the planar unit distance problem, first posed by Paul Erdős in 1946. OpenAI claims that the model autonomously cracked the problem. The real question is how much of this is marketing theatre and how much is AI genuinely creating new knowledge.

» Quote of the Week

4 years later, LinkedIn started to crackdown AI slop (allegedly). It only took 4 years!

Closing

With NVIDIA showcasing another record quarter, I don’t think anyone can argue that the AI “bubble” is something that may only exist in the imaginarium of some contrarians.

We may be overhyping what AI can do today, but there's still so much left to deploy that calling this a bubble assumes we've already seen the ceiling, and we haven't. The jump from infrastructure revenue to application revenue is the number to watch, it's the only metric that will tell us whether this ecosystem is creating real value or just recycling capital between labs and chip vendors.

See you next week!

— Ivo

When you browse social media, you notice a very typical behavior: you watch one video, and suddenly your timeline is flooded with more of the same. 5 years ago, it felt a bit like magic. But today, we talk about “the algorithm” as if it were a mysterious entity pulling strings in some Silicon Valley basement. The truth is much less dramatic, and much more interesting.

The algorithm isn’t inherently evil, it doesn’t sit there plotting your radicalisation. It’s just a chunk of code running cosine similarities and weighted averages, trying to predict what you’ll click on next. The trouble is what we interact with creates engagement. And the surest way to keep humans engaged turns out to be the worst way to keep them informed (rage-baits, fake news, or worse).

This post is about how recommendation engines work, why they tilt us toward echo chambers, and, because reading about a thing is never the same as seeing it, we’ll build one from scratch, point it at real news data, and watch the bubble form.

The Engagement Engine: How Recommenders Actually Work

A social media algorithm is, at its heart, a curator. Its job is to sift through millions of posts and serve you the ones you’re most likely to engage with: click, watch, like, share, rage-comment on. It does this based on one word: data.

Every action you take is a clue:

• Which posts you linger on (even without clicking)

• Which videos you watch, and for how long

• Which accounts you follow, mute, or block

• Which topics you search for at 1 a.m.

Using machine learning, the algorithm spots patterns in this firehose of behaviour. It’s constantly asking the same question: what keeps this person on the platform longer? Remember that this is the largest goal of any social media company: keeping you on the platform longer.

Two classic techniques sit underneath most recommender systems:

• Collaborative filtering finds users who behave like you and recommends what they liked. If Alice and Bob both loved The Matrix and Inception, and Alice also loved Interstellar, the system nudges Interstellar to Bob. Pretty easy to understand.

• Content-based filtering looks at the characteristics of what you’ve liked and finds similar things. If you watch a lot of cooking videos, it surfaces more videos tagged “cooking”, “recipe”, or “knife skills”, they resemble what you already enjoyed.

Real platforms blend these methods with hundreds of other signals. But the core idea is the same: learn from your behaviour, predict what else might grab you.

The algorithm doesn’t intend to show you bad or false content. It optimises for engagement. And one of the surest ways to keep humans engaged is to tap into our emotions, especially the strong, negative ones. Or videos of cats.

Building a News Recommender on Real Data

Let’s stop talking about this abstractly and build one. We will use real anonymised click logs from Microsoft News.

The dataset is called MIND (Microsoft News Dataset), published for academic research by Microsoft Research. This sample contains 50,000 users, over 51,000 English news articles across 17 categories (news, sports, finance, lifestyle, health, travel, and more), and 156,000+ real impression sessions, each recording what a user was shown and what they clicked on. The whole thing fits in about 30 lines of Python, although you don’t really need to now this:

import numpy as np
import pandas as pd
from scipy.sparse import csr_matrix
from sklearn.metrics.pairwise import cosine_similarity

# Build a sparse user × article matrix (1 = clicked, 0 = didn't)
matrix = csr_matrix((np.ones(len(clicks)), (user_rows, article_cols)),
                    shape=(n_users, n_articles))

def recommend(user_id, matrix, top_n=15, n_neighbors=50):
    """Find 50 most similar users and rank the articles
    they clicked that our user hasn't seen yet."""
    u = user_idx[user_id]
    
    # Cosine similarity between this user and everyone else
    sims = cosine_similarity(matrix[u], matrix).flatten()
    sims[u] = 0  # don't recommend to yourself
    
    # Take the top 50 most similar users
    top_neighbors = np.argsort(sims)[-n_neighbors:][::-1]
    weights = sims[top_neighbors]
    
    # Score articles by weighted sum of neighbour clicks
    scores = np.asarray(matrix[top_neighbors].T.dot(weights)).flatten()
    
    # Zero out articles the user already clicked
    scores[matrix[u].toarray().flatten() > 0] = 0
    
    # Return the top-scoring articles
    top_articles = np.argsort(scores)[-top_n:][::-1]
    return top_articles

That’s it.

Cosine similarity finds your fifty closest neighbours, people who click on the same kinds of articles you do.

We take the articles they clicked, weight them by how similar each neighbour is to you, and serve the top fifteen. This is the base of what powers a billion-dollar industry.

Coswhat similarity?

Cosine similarity might sound like something out of a math textbook, but bear with me, it's easier than it looks. To show you how it works, let's take a quick detour.

Imagine the following data points scattered across two axes, mechanical vs. biological, and cuteness:

Cosive similarity measures the angle between two arrows, each one starting from the origin (0,0) and pointing toward one of our data points.

The smaller the angle between them, the more similar the two items are. Think of it this way: if two arrows are almost pointing in the same direction, the items they represent share similar characteristics.

Take cats and dogs as an example. Both score high on ‘biological’ and high on ‘cuteness’, so their arrows point in nearly the same direction and cosine similarity returns a value close to 1 (its maximum).

But if we compare cats with teddy bears, although they are similar on the cute dimension, they are different on the biological axis:

But if we compare cats with teddy bears, although they are similar on the cute dimension, they are different on the biological axis, a cat is fully biological, while a teddy bear scores zero. This pulls their arrows apart.

The angle between them widens, and cosine similarity returns a lower value, reflecting that despite sharing one trait, these two objects occupy very different regions of our space.

And, of course, comparing cats to cars, give almost no similarity as the arrows between both point in different directions:

AI models use this kind of information to recommend content that is likely to trigger a similar response in you. Imagine a two-dimensional space where one axis captures how a video makes you feel (calm, entertained, outraged) and the other captures its topic. Every video gets plotted somewhere in that space.

If you click on a political video that makes you angry, and you watch it all the way through. The platform registers both dimensions: the topic and the emotional response. Using cosine similarity, it finds other videos whose ‘arrow’ points in the same direction (rage-baiting political videos) and serves them to you next. The more you engage, the more confidently the algorithm learns which corner of that space keeps you watching.

Meet User U92876 (let’s call it Joe): The Sports Fan

I picked a user from the MIND dataset whose reading history is pure sports, NFL power rankings, NBA trade rumours, MLB bans. It read twenty-five articles, all sport.

Let’s ask the recommender what to serve them:

The category breakdown:

• 40% sports

• 13% news

• 13% autos

• 34% scattering of everything else.

The algorithm recognises this person’s sports habit and feeds it back, but it also serves a reasonably varied diet. There’s politics, entertainment, lifestyle, finance. Not bad, right?

Now watch what happens.

The Moment of Curiosity

I simulated something far more common than a massive rabbit-hole binge: a moment of idle curiosity.

Our sports fan didn’t spend hours reading politics. Looking at their initial feed, they simply clicked on three items that caught their eye:

1. The news story about Joe Biden.

2. The news story about Mitch McConnell.

3. The video about Trump’s attacks.

Just three clicks in less than ten minutes of reading and watching. Three tiny breadcrumbs left for the algorithm and Joe goes on with his life during the rest of the day.

Now, if we ran those clicks through the basic 30 lines of Python code we wrote earlier, nothing much would happen. Mathematically, 25 historical sports clicks would completely overpower 3 new political clicks. The algorithm would still see a user who is 89% interested in sports, and the feed would barely budge.

But here is the very important secret sauce of modern social media: Recency Weighting (or Time Decay).

Real algorithms don’t treat all your clicks equally as a click you made three years ago is practically ancient history; a click you made three minutes ago is gold. To keep you hooked in the current session, platforms apply a heavy multiplier to whatever you are doing recently.

A single line of code implements this in the algorithm we saw earlier. If we decide that the most recent clicks should carry up to 100 times more weight than older ones, we could write something like this:

time_decay_weights = np.array([0.1 if historical_click else 10.0 for click in user_history])

If we do this, let’s run the recommendations again:

Here’s the damage that just 3 clicks have done to our Time Weighting recommendation system:

Political news went from 13% to 40% of the feed. A 3x increase. From one evening of clicking and reading three pieces of news.

Sports (the thing this person has read for years) can drop from the dominant category to second place. The algorithm didn’t pause to think “hold on, this person has 25 sports articles in their history, and one evening of politics doesn’t define them.”

It doesn’t think, it just recalculates time weighted similarity matrices, found a new set of neighbours and served what other users that clicked on this may enjoy.

Two things jump out:

• The speed. It may take one evening to flip a user’s entire feed composition. Real platforms recalculate faster than this demo as they update in real time. You probably notice this in your feed with ads related to products you’ve been searching lately.

• What disappears. This isn’t just about what the algorithm adds, but also about what it removes. The user’s informational diet didn’t just get more political, it got narrower. And narrower is the real danger here

Note: real platforms don't publish their decay constants, so this is illustrative, not a measurement, but the mechanism is real and the direction is what matters. My 100x example is possibly an exaggeration of the recency bias.

What research tells us

You now know how clicks influence what the math of what the algorithm shows you next.

But this gets worse - content that makes us angry, fearful, or shocked glues us to the screen far better than content that makes us feel good or informed. Social media companies didn’t engineer this consciously, their algorithms simply discovered it.

A massive 2025 study analyzing the digital trace data of 25,000 SmartNews users found that humans possess a trait-level “negativity bias” when selecting news. Evolutionarily, we are hardwired to pay attention to threats, avoiding danger was critical to our ancestors’ survival. What happens when this ancient instinct meets modern machine learning? The study confirmed that personalized recommendation feeds take our inherent negativity bias and actively augment it.

Furthermore, data from researchers analyzing hundreds of millions of posts on platforms like Facebook and X (formerly Twitter) reveals that social media users are roughly 1.91 times more likely to share negative news links than positive ones. Negativity equals virality, and the outrage loop is born.

The Cognitive Toll: It’s Not Just What You Think, It’s How You Think

The impact of these algorithmic loops isn’t just about the type of content we consume; it’s about how it fundamentally alters our brains. A recent 2025 systematic review analyzing 71 studies and 98,299 participants of short-form video feeds (like TikTok, Instagram Reels, and YouTube Shorts) found profound cognitive consequences.

Increased engagement with these endless-scroll platforms is associated with poorer cognitive performance, specifically impacting our sustained attention and inhibitory control. Psychologists point to a dual process of habituation and sensitization to explain this phenomenon. The rapid, high-stimulation nature of short videos desensitizes us to slower, more effortful tasks like reading a book or deep problem-solving. At the same time, the algorithm’s instant delivery of curated content sensitizes our brain’s reward system, reinforcing impulsive engagement patterns and encouraging the habitual seeking of instant gratification.

Heavy users of these platforms exhibit reduced electrophysiological activity during attention-demanding tasks. Some researchers even point to structural differences in key cognitive control regions, including the prefrontal cortex and striatal reward circuits, linked to this constant bombardment of highly rewarding algorithmic stimuli.

The Societal Cost

Due to mathematical matrices, each of us in our own personalised bubble of information.

In the short term, it’s annoying for most people. But zoom out and the picture darkens. When algorithms feed us content that confirms what we already believe, we experience confirmation bias on steroids.

These filter bubbles deepen the social divide we have today. We will continue to be extremely divided and there is no end in sight for this chasm.

Misinformation thrives in closed loops because false stories don’t get exposed to scrutiny outside the bubble. By the time a fact-check goes out, the original lie has done a lap around the platform and built a small army of believers.

And democracy, which depends on a shared baseline of reality and some willingness to argue in public, takes a hit when citizens occupy entirely different reality bubbles.

Reclaiming Your Feed

You’re not powerless here. The algorithm is responsive but there a couple of things you can do, that, although annoying, may take you out of your bubble.

The same mechanism that built your bubble can be used to widen it. Some practical moves:

• Diversify the inputs. Actively follow a few sources outside your comfort zone. If you lean one way politically, follow some thoughtful voices from the other side.

• Reset periodically. Clear your watch history. Use “Not Interested” on suggestions that keep haunting you. Try the platform logged out, or in incognito, and notice how different the world looks without your data.

• Use chronological feeds. Most platforms still let you switch off the algorithmic ranking and just see posts from people you follow, in order.

• Pause before sharing. Every like, comment, and share is a vote for “more of this, please.” If something makes you furious, that’s exactly when the algorithm is most likely to be exploiting you.

• Limit the time. Set screen-time limits. Schedule offline hours. The less you depend on the feed for your information diet, the less power it has to shape what you believe.

Beyond the Bubble

Hope this blog post informed you on how these recomendation systems bubbles work. We built a recommender on real news data, and it took three clicks to flip a sports fan’s feed from 40% sports to 53% politics.

The first step to breaking free is simply being aware. Next time you find yourself in an online frenzy, take a breath and ask: Why am I seeing this? Who benefits from me reacting this way? The answer usually traces back to an algorithm doing its job, and that job is rarely “informing you.”

Stay informed, stay open-minded, and happy exploring on your data journey.

— Ivo

This week, Cerebras, the AI chip company, raised $5.5 billion in its IPO, with an order book 20x oversubscribed - it’s another sign that the infrastructure layer of AI is going to continue heating up for the next few months.

On OpenAI’s side, the company launched a personal finance product that wants access to your bank account, and reportedly preparing legal action against Apple, its own partner.

Meanwhile, Anthropic is quietly going downstream with Claude for Small Business, dropping into QuickBooks, HubSpot, and DocuSign. Salesforce is spending $300 million on Anthropic tokens this year, and Cisco just cut 4,000 jobs to fund its own AI pivot while reporting record revenue in the same report.

Let’s get into it.

ps: One more thing: the OpenAI vs. Musk trial reached a verdict yesterday: Musk lost. I'll do a proper deep dive next week, but I wanted to flag it, even though it belongs to week 21.

» Stat of the Week: Cerebras prices at $185/share, raising $5.55B

The AI chip maker’s IPO closed with a market cap near $56 billion on day one, the largest U.S. tech IPO in years.

The order book was 20x oversubscribed. For context: the company was valued at $8 billion just eight months ago.

» Industry Highlights: Market & Product

Vapi hit a $500M valuation after beating 40 rivals to win Amazon Ring’s AI voice platform contract. Its enterprise business has reportedly grown 10x since early 2025 as companies move customer support and sales calls to AI agents.1

OpenAI launched a personal finance feature inside ChatGPT that lets users connect bank accounts directly.2 The dashboard shows portfolio performance, spending, subscriptions, and upcoming bills. The product assumes people trust OpenAI with their bank account. That's a big assumption.

In another busy week for the company, it has enlisted an outside law firm to explore legal options against Apple, due to a potential breach of contract3. This is nothing new, as it wouldn’t be the first partner to feel burned by Apple. The AI companies are learning what media companies, app developers, and Spotify already figured out: being inside Apple’s ecosystem and being protected by it are very different things.

On a quieter week on Anthropic’s side, the company launched Claude for Small Business4, a one-toggle install that drops Claude into QuickBooks, PayPal, HubSpot, Canva, DocuSign, and Google/Microsoft 365. Small businesses make up 44% of U.S. GDP but have been largely left out of the enterprise AI wave - Anthropic makes yet another play for a new market.

Leaving the model and product layer, and entering infrastructure. Cerebras, the AI chip maker officially landed on the Nasdaq under ticker CBRS5, raising $5.55B at $185/share, above its already-raised expected range. Cerebras designs its Wafer Scale Engine chips specifically for AI inference, essentially, the compute required to run models rather than train them.

And finally, two stories from yesterday's tech darlings, now firmly in incumbent territory:

• Salesforce expects to spend $300 million on Anthropic tokens this year, and CEO Marc Benioff is already teasing AI coding natively inside Slack.6

• In another round of tech layoffs, Cisco cut nearly 4,000 jobs while simultaneously reporting record quarterly revenue7. The official reason is to free up budget for AI investment.

» Industry Highlights: Research & Science

NVIDIA’s research lab (NVLabs) dropped SANA-WM this week: a 2.6B parameter open-source world model that takes a single image and a camera trajectory as input, then generates a 60-second, 720p video from it8.

Check a simulation of a 3D world generated by the model here:

» Economic Highlights: The World of AI

On the geopolitical side, US Treasury Secretary Scott Bessent announced that the two countries will begin formal AI safety talks, focused specifically on preventing advanced models from falling into the hands of non-state actors.9

On a society level, Canva dropped a new report of AI in Marketing10: 97% of marketers use AI daily, while 78% of consumers still prefer human-made content. There's a growing chasm between the people deploying these tools and the people on the receiving end, and mentions of 'AI slop' across social media have risen ninefold in a year. This chasm will be a society problem and there’s a growing rumble between AI believers and haters.

Thank you for reading and for supporting the newsletter. I’m always interested in getting your feedback on the newsletter, so feel free to add your thoughts here.

— Ivo

Some of you may know, but a couple of months ago I wrote a book. It’s called The AI Handbook for the Curious Mind, and it’s available on Amazon or Gumroad. As this newsletter as been gaining some traction laterly, I’ve decided to do something a little different for paid subscribers here on The Data Journey.

Every two weeks, I’ll be releasing a full chapter here, exclusively for paid subscribers. I keep looking for ways to give back to the people who support this newsletter, and sharing the book felt like one of the more meaningful ones. 😄

If you want to check other chapters, check:

• Chapter 1: Introduction

• Chapter 2: Artificial Intelligence 101

• Chapter 3.1: Statistics Intro

• Chapter 3.2: Data Types

If you’ve been following along, you’ve now survived two chapters on statistics, and you know the difference between continuous and categorical data. You’ve also met distributions, means, medians, and the sneaky ways averages can mislead you.

Now we get to the part everyone came for. This is the chapter where machine learning stops being a buzzword and starts being something you can picture.

If you try to guess a house’s price using just the average price in the neighbourhood, you’ll miss everything that makes one house worth more than another: its size, its condition, its age, the number of parking spaces. You’ll be wrong a lot, and in consistent, expensive ways.

This is exactly where machine learning will help you.

Read more

This week, the OpenAI courtroom drama kept delivering, with Greg Brockman own journals playing against him. Not the first time a personal journal has walked into a courtroom, but probably the first time one showed up to decide a high tech battle.

In other news, we AI-and-jobs debate continues to explode: Cloudflare cut 1,100 people the same quarter it hit record revenue, and Airbnb disclosed that 60% of its new code is now written by AI.

But, due to the incresing public backlash, big tech is walking back the doom-and-gloom labor narrative (starting with Jensen Huang). Whether that is a true change of heart or just better PR instincts is left as an exercise for the reader.

Let’s dive in into this weeks’ digest.

» Stat of the Week: 60%

Airbnb CEO Brian Chesky disclosed this week that 60% of all new code shipped in Q1 2026 was AI-generated.1

» Industry Highlights: Market & Product

In a turn of events, Anthropic signed a compute deal with SpaceX, gaining access to over 220,000 NVIDIA GPUs and 300 megawatts of new capacity at Colossus 1. The immediate consequence is that the company doubled rate limits for Claude Code (Pro, Max, Team, and Enterprise plans) and significantly raised API rate limits for Opus models.2 Remember that Elon Musk criticized Dario Amodei and Anthropic several times in the past, but it seems that a common enemy made them bury the hatchet and work together. This also raises the question if xAI (that belongs to Elon Musk) is aiming to compete more on the infrastructure layers than on the model layer.3

On the other side, OpenAI launched GPT-5.5 Instant as the new default model for ChatGPT, replacing GPT-5.3 Instant. The model improves hallucination rates in sensitive domains like law and medicine, scores 81.2 on the AIME 2025 math benchmark (up from 65.4), and features deeper context management.4

Still on OpenAI’s models, openRouter crunched the real cost of GPT-5.5: users are paying 49–92% more than with 5.4. The model is less wordy, but not nearly less wordy enough to offset the 2x price increase on tokens. The efficiency gains across the models is being collected by OpenAI.5

Airbnb revealed that 60% of Q1 2026 code was AI-written, and its customer support bot now resolves 40% of issues without human escalation. But CEO Brian Chesky also offered a take on AI’s limits when it comes to Chatbots: these AI systems don’t deal well with photos, and can’t perform direct manipulation (you have to type everything rather than adjust sliders) - this is probably changing with Agentic native systems such as Claude Cowork.6

Doing a quick side-trip to the other continent, DeepSeek is in talks to raise its first-ever external funding round, with its valuation reportedly jumping from $20 billion to $45 billion within weeks. The round is expected to be led by a Chinese state investment vehicle (surprise /s). Worth watching this first raise from a Chinese company.7

Finally, the second week of the Musk v. Altman trial delivered something taken out of a soap opera: a tech executive being cross-examined using his own diary. Greg Brockman took the stand, with personal journal entries asking “Financially, what will take me to $1B?” The judge had already cited those entries during the trial, with Musk stating that this is significant proof that the OpenAI original mission (non-profit) was a lie.8

» Industry Highlights: Research & Science

A Chinese court ruled this week that AI-driven automation is not a legally valid reason to fire employees (with a nuance I’ll explain in the next paragraph).9

Most outlets ran this as 'China protects workers from AI.' The actual story is more boring, but also more interesting. Chinese labor law has two flavors of termination: one cheaper, one expensive. Neither is illegal. and the company fired the worker using the cheap version, arguing that AI adoption counted as an 'objective major change in the business environment.' The judge disagreed, AI is a management call, not an act of God, and ordered the full severance. Had the company simply paid the higher amount from day one, this would probably not made the news.

Regardless, China is betting on AI dominance while also building legal protections for workers affected by it. Whether that’s sustainable, or just a delay, is an open question. But, undeniably, China seems to be approaching this transition with more structural intent than the West - pushing hard on AI development while simultaneously building legal guardrails for workers, which will make the country more prepared for the disruption of the next 10 years.

On the west, Jensen Huang at the Milken Institute made the case that AI is “creating an enormous number of jobs,” arguing that automating a task does not eliminate the broader function a worker serves. He also pushed back against “doomer” rhetoric, warning that overhyping AI’s destructive potential could make it unpopular enough that Americans don’t engage with it. The man who sells the shovels in a gold rush has a natural interest in making the gold rush sound good, but expect a large pivot by the big tech players as a way to prevent the enormous backlash they will be suffering by displaced or fearful workers.10

» Economic Highlights: The World of AI

Cloudflare cut 1,100 employees in its first mass layoff ever, even as quarterly revenue hit a record $639.8 million (up 34% year-over-year). The company’s internal AI usage grew 600% in three monthsa and the rationale was that highly productive AI-powered employees simply need fewer support staff around them.11

On the same side of the coin, Coinbase announced layoffs due to the same reason: the crypto exchange joins a growing list of companies using AI as justification for headcount reductions even during periods of strong performance.12

💬 Quote of the Week

Closing

If this week had a tagline it would be: the numbers are great, have you considered updating your CV? Cloudflare posts record revenue and cuts 20% of its workforce in the same breath. Coinbase does the same. And the CEO of the company that makes the chips running all of this tells the Milken Institute that AI is a job-creation machine. Which, to be fair, it may be and only time will tell.

The OpenAI vs. Musk trial continues. The layoffs continue. The models get better and cheaper and the token prices go up anyway. Despite all of this, unemployment rate keeps relatively stable in the US (although it may be rising silently).

See you next week.

— Ivo

It’s called LeWorldModel (LeWM), and it comes from a team at University of Montreal, NYU, Samsung Sail, and Brown University, including Yann LeCun, Meta’s former Chief AI Scientist and arguably one of the most prominent skeptic of the “LLM” school of thought.

Before I explain what the paper does, let me set the stage.

The LLM orthodoxy and its discontents

If you’ve been following AI for the past few years, you’ve been swimming in Large Language Models. Products such as ChatGPT, Claude or Gemini are all large language models. They work by predicting the next token in a sequence, trained on enormous amounts of text. They’re extraordinarily capable at language tasks, and increasingly good at reasoning, coding, and multimodal understanding. These products proved something remarkable: most of our world is, at its core, encodable in language.

LeCun has been publicly skeptical of this approach for years. His argument is that LLMs learn statistical patterns over language tokens. They don’t have a model of the world. They can’t plan. They hallucinate confidently because they’re pattern-matching machines with no grounding in physical reality.

His alternative for the AGI problem lies in world models. The idea is that a truly intelligent system should be able to build an internal representation of how the world works, and use that representation to reason about what will happen next before taking any action. Like a chess player visualizing 10 moves ahead, except for robotics and physical tasks.

This sounds great in theory. In practice, it’s been extremely hard to make work.

The collapse problem

So what’s the main problem of world models? Representation collapse.

Imagine you’re training a model to predict the next video frame. To learn well, it needs to build meaningful representations of what’s in each frame: positions, objects, movement. But there’s a shortcut as models discover that if they map every frame to the exact same internal representation (a flat, featureless blob) then every prediction looks the same, and the prediction error is technically minimized.

This is like a student who, faced with a hard exam, decides every answer is “C.” Technically minimizes wrong answers if the exam is graded on a curve. Completely useless in practice.

Existing methods to avoid collapse are... messy. They involve exponential moving averages, stop-gradient tricks, pretrained encoders borrowed from other tasks, or loss functions with six or seven terms that need careful balancing. The PLDM method, which LeWM is compared against, has seven loss terms and six hyperparameters to tune. Getting them to cooperate is a part-time job.

What LeWorldModel actually does

LeWM’s contribution is cool: it trains a world model end-to-end from raw pixels with only two loss terms and one effective hyperparameter.

The architecture has two components:

The encoder takes a raw video frame and compresses it into a compact latent representation, a small vector that captures what’s happening in the scene. It’s the model’s “mental image” of the world, stripped of irrelevant visual noise.

The predictor takes that latent representation plus an action and predicts what the next latent state should be. Given “robot arm at position X” and “move 5cm right,” predict the new latent state.

The two loss terms are:

1. Prediction loss: how wrong was the predicted next state compared to the actual next state?

2. SIGReg: a regularizer that forces the latent space to follow a Gaussian distribution, making collapse mathematically impossible. Instead of heuristics, this is a proper statistical test applied along random projections of the embedding space.

In technical terms, you don’t need a bunch of things such as frozen pretrained encoders, no stop-gradient or no seven-term loss balancing act.

The result is not only relevant on the accuracy level, but on the scalability: LeWM plans 48× faster than foundation-model-based world models, trains stably, and performs competitively across 2D and 3D control tasks, such as manipulation, navigation, locomotion.

Ok, but how is this different from an LLM?

Let me make this concrete, because this is the part I find most intellectually interesting.

An LLM works like this:

“The cat sat on the ___” → “mat” (next token prediction, over text)

A world model works like this:

“Robot arm at position X, action: move left 5cm” → predicted next state (next state prediction, over latent representations of the physical world)

The fundamental difference is what’s being predicted and in what space.

Read more

Some of you may know, but a couple of months ago I wrote a book. It’s called The AI Handbook for the Curious Mind, and it’s available on Amazon or Gumroad. As this newsletter as been gaining some traction laterly, I’ve decided to do something a little different for paid subscribers here on The Data Journey.

Every two weeks, I’ll be releasing a full chapter here, exclusively for paid subscribers. I keep looking for ways to give back to the people who support this newsletter, and sharing the book felt like one of the more meaningful ones. 😄

If you want to check other chapters, check:

• Chapter 1: Introduction

• Chapter 2: Artificial Intelligence 101

• Chapter 3.1: Statistics Intro
  

Last time, we spent a while with continuous data, the kind that flows smoothly across a range such as house prices, heights and temperatures. We talked about means, medians, standard deviations, and how the shape of a distribution can already tell you something interesting before you’ve run a single AI model.

But data doesn’t come in just one flavour. Most real-world datasets are a messy mix of different types, and knowing which type you’re looking at changes everything about how you treat it.

So this week: the other kinds.

Read more

The drama is real for this week: the Musk vs. OpenAI court hearings began, while Microsoft and the company also clarified their relationship for the general public.

On the Anthropic side, the $900 billion valuation rumour continues, with even a Bay Area home being listed in exchange for Anthropic shares. Talk about hype!

Grab your coffee and let’s dive into week 18,

» Stat of the Week: Musk’s Donation to OpenAI

» Industry Highlights: Market & Product

Anthropic, eyeing a $900B valuation, received a fresh 5B investment from Amazon (bringing its total to $13B)1. As usual, in the circular nature of these deals, Anthropic then committed to over $100B of Amazon Web Services spend over ten years. The structure echoes Amazon’s earlier OpenAI deal almost line-for-line: invest dollars, recapture them as cloud revenue.

The company is so hot right now, that it’s even having repercussions in the housing market. An investment banker put a 13-acre Mill Valley property on LinkedIn asking to swap it for Anthropic shares2, with the buyer keeping 20% of the upside through the lockup. He described it as a “diversification play” and it’s funny on the surface, but it’s also a quiet signal that private AI shares are now treated as harder currency as Bay Area land. Sign of the times?

On the other side of town, Microsoft & OpenAI announced a new phase on their partnership. The two companies amended their agreement on April 27. The most significant update: OpenAI can now ship across any cloud (Azure stays the primary)3, Microsoft keeps an IP licence through 2032 but it’s no longer exclusive, and the AGI clauses have been simplified into dates and caps. Less “what if we invent the future together,” more “here’s who pays whom and until when.” The most famous AI partnership in tech is now a normal commercial deal. This comes as little surprise, as the companies have been signalling a significant loosening of their exclusive partnership agreement for some time.

Still on Microsoft, Github Copilot is moving to usage-billed pricing4, a similar move done by other AI products and companies. Pro stays $10/mo with $10 of credits, Pro+ $39 with $39, Business $19/seat with $19. Code completions remain free, but the fallback-to-cheaper-model safety net disappears.

Finally, Meta is training agents on its own employees’ keystrokes and mouse movements. Meta says agents need real examples of how people actually use computers, clicks, dropdowns, navigation. Internal tool, certain applications, “safeguards in place.” The privacy implications, especially as this pattern spreads to consumer products, are only starting to surface. Imagine if your operating system used your keystrokes to train AI models?5

» Industry Highlights: Research & Science

David Silver, the DeepMind researcher behind AlphaZero, raised $1.1B at a $5.1B valuation (Sequoia, Lightspeed, Nvidia, Google, the UK’s Sovereign AI fund) for Ineffable Intelligence. The pitch is a “superlearner” that discovers knowledge through reinforcement learning alone with no human-generated data. Similar to World Models’ startups by Yann LeCunn and Fei-Fei Li, this is yet another startup betting on a future with techniques that will outperform Large Language Models. The company aims to use Reinforcement Learning to achieve AGI.6

Finally, we have confirmation of what we’ve been suspecting for a long-time: AI labs use each others’ models to improve theirs. Under oath, Musk said xAI “partly” distills OpenAI models to train Grok, framing it as standard industry practice. The technical detail matters less than the admission: the same labs that have spent two years framing distillation as a Chinese problem just had a US frontier lab confirm, in court, that they do it too.7

» Economic Highlights: The World of AI

On a twist that surprised absolutely no one, China blocks Meta’s $2B Manus acquisition.8 The NDRC (National Development and Reform Commission) formally ordered Meta to unwind the deal, four months after announcement, another showcase of the AI wars between the two largest world empires.

» Track of the Week: Musk v. Altman, Day One

The Musk v. Altman trial opened in Oakland on April 27th and it delivered immediately.

Musk’s argument is that OpenAI was built as a charity for the public good, and converting it into an $850B for-profit machine is, in his words, looting. “If it’s okay to loot a charity, the entire foundation of charitable giving will be destroyed.” OpenAI’s counter is also blunt: Musk left because he didn’t get control, launched a competing for-profit AI company a year later, and is now using the courts to slow down a rival he couldn’t beat in the market.

Grab the popcorn!

Closing

This week was not very hot on new models, but it was hot on the power plays between big tech: Amazon recaptures its Anthropic investment as cloud spend. Microsoft decouples from OpenAI, xAI distills OpenAI models and Musk faces OpenAI in court.

As usual, the boundaries between competitors are blurrier - are these companies competing or partnering with each other?

What's clear is the bigger game: the US and China are in a full-blown race for AI dominance, and every deal and acquisition block is just a move on that board.

— Ivo

Some of you may know, but a couple of months ago I wrote a book. It’s called The AI Handbook for the Curious Mind, and it’s available on Amazon or Gumroad. As this newsletter as been gaining some traction laterly, I’ve decided to do something a little different for paid subscribers here on The Data Journey.

Every two weeks, I’ll be releasing a full chapter here, exclusively for paid subscribers. I keep looking for ways to give back to the people who support this newsletter, and sharing the book felt like one of the more meaningful ones. 😄

Read more

You’re not imagining it. I’ve covered this in the blog before, when I went through Anthropic’s report on AI’s labor market impact: junior-level roles in AI-exposed occupations are showing a real, statistically significant drop in entry rates for workers aged 22–25. People aren’t getting too laid off (although we have layoffs from big techs, the unemployment rate hasn't moved significantly), they’re just not getting hired in the first place.

During the past year, I’ve interviewed more than 500 candidates for DareData, but also recommended dozens of candidates for companies that reach out to me asking for Data Scientists, ML Engineers and Product Managers. In this blog post, I want to share what’s actually working for the people who break through. None of this is “hustle harder” advice and it’s not technical, for sure. And a lot of what hiring managers care about isn’t on the list of things candidates think they should be optimising for - nor in the job post.

Let’s get into it.

A note: this one's mostly for students, so I'm leaving it free. No paywall on advice that should reach the people who need it most. Probably, students don’t have resources to pay for the paid version of this newsletter, so I’m leaving it open.

1. Be the person who takes care of things

This is, by some distance, the most underrated skill in the modern job market. It’s also the one I look for first when I’m interviewing someone for a junior role - especially looking for situations where they’ve assumed responsibility when they didn’t need to. I’ve spoken about this skill over and over on my leadership blog.

“Taking care of things” sounds vague, but it’s simple: when something is on your plate, everyone knows that you will find the resources to get it done (note that this doesn’t mean that you have the resource, but that you will find the resources needed to complete a task).

If you’ve ever worked in a team, you know exactly the kind of person I mean, and you also know how rare they are.

The reason this skill is so valuable now is because AI handles the task layer fairly well. What it can’t do is own a thread of work end-to-end across humans, systems, and ambiguity. That’s the gap that’s getting more valuable and if you become known for closing loops, you become hireable in a way that doesn’t depend on which framework is hot this year.

You can practise this skill anywhere: at school, in volunteer work, even at home. Take on the task that looks too big and just ace it through.

2. Learn to disagree without being a pain

The cliché version of teamwork advice is “be a team player,” which is too vague. The thing I actually screen for in interviews is whether someone can disagree with me constructively in a 45-minute conversation.

I’ll float an opinion that’s deliberately a bit off, about an architectural choice, or a process question, or how to scope a project. I want to see how the candidates think and if they can trade ideas and opinions without becoming defensive. The bad responses are the obvious ones (just agreeing with me, or arguing aggressively).

Disagreeing well is a skill that compounds with experience, but you can get a head start just by watching how badly most young people do it. :-)

3. Volunteer somewhere

Volunteering is the holy grail of networking.

My first proper gig with DareData came through a volunteer organisation. I wasn’t applying for jobs at that point, I was just helping run things at a non-profit context where I happened to meet people who later thought of me when a leadership position opened up. Volunteering in spaces tied to the work you want to do is how you expose yourself to luck.

The mistake juniors make is treating volunteer work as a CV line. The CV line is the byproduct, while the actual value is that you spend time around people who do things, and those people remember you. Six months later when somebody says “we need someone for X”, your name is on top of the list, especially if you follow advice number 1.

If you’re early career, find a student club, an NGO, an open-source project, a meetup group. To be useful in a place where useful people are paying attention.

4. Your portfolio is your resume now

If you’re a technical person, GitHub matters and a personal website matters. Anything that lets a hiring manager see your work matters in a world flooded by AI generated CVs.

When I’m reviewing a junior application, the CV tells me what you claim but the portfolio tells me what’s actually true. As I mostly hire AI engineers, I can tell within 30 seconds of looking at someone’s GitHub whether they understand what they’re doing, the commit messages, the README quality, the structure of the projects, whether the repos are abandoned shells or actual working things. Yes, and I can spot AI generated slop in repos too!

You don’t need impressive projects, you just need real projects tied with something you enjoy and love. The size doesn’t matter, but how much passion you put in it, does.

If you’re not technical: the same logic applies, just in a different format. A portfolio site with case studies, a few well-written analyses on Medium, A presentation deck from a real project you ran. Anything that lets someone evaluate the work, not just the claim.

5. Write in public

Most young people think they don't have anything worth saying until they're more experienced and that's wrong. I've read pieces from students so full of curiosity that I'd happily read a long-form essay from them.

Pick a topic you care about and start writing about it publicly. Substack, Medium, LinkedIn, your own blog, it doesn’t matter. The platform matters less than the consistency. The reason this works is simple: most juniors are invisible to hiring managers until they apply. If you write publicly about your field for six months, you arrive at the interview already half-known. The hiring manager may have read your stuff or noted you.

The trick is to write about what you’re learning, not what you’ve mastered. It’s a win-win: you may get noticed, but also improve on the compreheension of topics you want to know about.

Oh, and never let AI define your writing style. :-) Everyone can spot obvious AI prose from a mile away now, don't cut that corner.

6. Get fluent at working with AI before AI works without you

This is probably the most obvious advice.

Every junior I interview now, I quietly assess for one thing: can they work with an AI tool intelligently? Today, working with AI intelligently is not copy-pasting code or paragraphs from AI tools. I mean more: do they know when to trust the output, when to push back, when to verify, when to throw it out? Do they treat the model like a teammate they’re supervising, or like an oracle?

The juniors who are getting hired right now treat AI as a multiplier on their judgement. The ones who aren’t, treat it as a substitute for their judgement.

Do real work with these tools, frequently, and pay attention to where they help you and where they hurt you.

If you’re a student or a recent grad reading this, I don’t want to fool you: the market is harder than it was, and the anxiety you’re feeling is rational. Pretending otherwise would be wrong.

But the things that get you hired right now are not the things AI is automating. AI is good at tasks, but it’s not good at owning a thread of work across humans and ambiguity, at disagreeing constructively in a room full of opinions, at noticing what nobody assigned, at being the person colleagues trust. That’s still you.

Humans aren’t task executors. We’re the layer that maps how tasks connect, who needs what, when something is going off the rails, and what’s actually worth doing in the first place. That layer is getting more valuable, because there’s more output flying around that needs someone with judgement to make sense of it.

Show up as the kind of candidate who already looks like they’re doing the job, and the door opens. When you establish yourself as a reliable future professional, your reputation does the networking for you, no applications required.

This week alone I’ve highlighted 20 events worth mentioning on our journey of understand how AI is moving. I continue to feel the space heating up every week and showing no signs of stopping.

This week we will speak of new launches by the Big Labs, and the escape of Claude Mythos from its lab. In this past week there were also a lot of products released from Anthropic (Claude Design), OpenAI (GPT Agent Framework) and Google. I’m feeling the big labs want to be players (model providers) and referees (all-round prouct providers) in the same game, and I don’t think they will be able to win every battle.

On another note, reports that Anthropic may remove Claude Code from its $20 Pro plan reinforce the theory that major labs are currently subsidizing market share to capture as much demand as possible.

We’ll approach all these topics in this weeks’ digest, so, buckle up!

Stat of the Week: $100 billion

Anthropic pledged to spend $100 billion on AWS over the next 10 years in exchange for a fresh $5B investment from Amazon. To put that in perspective: that’s roughly 3x Anthropic’s current annualized revenue committed, in advance, to a single cloud vendor.1

Industry Highlights: Market & Product

Claude Opus 4.7: Anthropic’s new flagship lands with what looks like a genuine step change on SWE-bench Pro (64.3% vs GPT-5.4’s 57.7%)2, a third of the tool errors of Opus 4.6. Same price as 4.6, but, probably, Anthropic will explore with changing their pricing in other creative ways.

The company has begun flirting with the limits of the 'all-you-can-eat' subscription model, recently launching a '2% test' that gates Claude Code behind a higher paywall for new users. This experiment suggests that the extreme compute costs of agentic coding are forcing a shift away from subsidized $20-a-month tiers toward more sustainable, usage-based pricing. One day the “capture-the-market” stance by labs will be over.

Claude Design + Canva: The company has also launched Claude Design3, a visual collaboration surface powered by Opus 4.7, in deep partnership with Canva. Describe what you want, Claude builds it, and you can export to PPTX, PDF, or straight into Canva’s editor. It also reads your codebase to pick up your brand system automatically. I’ve used the tool this week to build a presentation and it was impressive:

Beyond the wow moment, we’re likely seeing the birth of a new 'design slop' aesthetic: an over-reliance on faux-intellectual italics and rigid, 'boxed' layouts.

And finally, Bloomberg reported that a Discord group obtained access to Anthropic’s restricted cyber model Mythos through a third-party vendor.4 They used an employee’s credentials at a contractor and have apparently been using the model regularly. Anthropic confirmed it’s investigating. Remember that the entire point of Project Glasswing was to keep Mythos out of “the wrong hands” via a whitelist of 11 organizations.

On the other side of the valley, OpenAI has witnessed a triple exit: CPO-turned-Science-lead Kevin Weil, Sora head Bill Peebles, and enterprise CTO Srinivas Narayanan all left on the same Friday5. Only 2 of 11 OpenAI co-founders remain.

At the same time, the company is firing on all cylinders with new product launches. First, they launched a life-sciences specialist model, named after Rosalind Franklin, built for drug discovery, genomics, and protein reasoning.6 Read more on OpenAI.

But OpenAI didn’t stop here, they’ve launched GPT-5.5 & OpenAI Workspace Agents. OpenAI dropped the GPT-5.5 System Card7, positioning the model for “complex, real-world work.” Accompanying this is Workspace Agents8, a major pivot from “GPTs” to shared, autonomous team members for ChatGPT Business/Enterprise. They can connect to Slack and internal data to handle end-to-end workflows as the company tries to compensate for the time they’ve lost on capturing the enterprise market.

Not to be outdone, and almost simultaneously, Google announced the launch of the Gemini Enterprise Agent Platform, effectively replacing the Vertex AI roadmap. It treats agents as managed workloads with full DevOps lifecycles. They also teased TPU 8t and 8i, promising 3x better performance for the next generation of training.9

The Chinese lab DeepSeek continues to haunt the “Big Three” by releasing DeepSeek-V4-Pro (1.6T parameters)10. It allegedly closes the gap with frontier models like GPT-5.4 at a fraction of the cost ($1.74/M tokens input), proving that the “moat” of high-end performance may be narrower than it looks.

Finally, Cerebras files for IPO, again11. The wafer-scale chip startup filed S-1 after withdrawing in 2024 over the G42 review. $510M in 2025 revenue, a recent $23B valuation, an AWS deal, and a reported $10B+ OpenAI contract.

🔬 Industry Highlights: Research & Science

Printed neurons that talk to real ones: Northwestern’s Mark Hersam group published in Nature Nanotechnology a flexible, printable artificial neuron that produces signaling patterns close enough to biological spikes. In a sentence: we may have successfully artificial neurons able to communicate with biological neurons.12

🏛️ Economics and Politics Highlights

In another round of potential layoffs, Microsoft is reportedly offering voluntary buyouts to 7% of its U.S. staff, signaling a massive internal reallocation of capital toward AI.

Meta seems to be going the same way, with 20k expected layoffs occurring in the next few months.13

With agentic AI, are companies considering that “size = disadvantage”?

💬 Quote of the Week

The labs are playing chess while acting like they’re playing checkers. They’re positioning themselves as both the vendors and the referees (building the models, the platforms, the design tools, the agents) while simultaneously claiming they’re just “democratizing AI.”

But markets don’t work that way. You can’t be the manufacturer, the distributor, and the product builder. Something will got to give.

— Ivo

The 2026 edition landed last week and it’s a mamooth of a doc with 423 pages. I read through it so you don’t have to.

Let’s check the best highlights!

Read more

These past few weeks, we’ve witnessed our first major signs of political unrest fueled by AI. Sam Altman’s home was targeted twice by protesters, individuals clearly angered by the rapid trajectory of this technology. While civil unrest isn’t exactly a novelty in the US (a country that has seen its share of upheaval, including the assassination of two presidents), it is a telltale sign what we should expect. I’ve been sounding the alarm on this for nearly a year: we are entering a period of massive disruption, and unfortunately, the violence has just started.

Beyond the headlines, the world keeps turning. We’ve seen steady technical advances and the release of a significant report from Stanford University. It’s a massive document, and I’ll be dropping an essay later this week to detail and analyze all 400 pages of it for you.

Oh, and keep your guard up, the Claude Mythos model is beginning to circulate, and it’s already posing a credible threat to computer network infrastructures as we speak.

Buckle up and grab your coffee. Welcome to week 16 of 2026.

📈 Stat of the Week: 69 vs. 21

We see a fundamental trust issue reflected in how different groups view the economic impact of AI. According to data from the Stanford 2026 AI Index Report, there is a divide in expectations about AI:

• Experts (69%) are highly optimistic that AI will benefit the economy, while only 21% of the general public shares this view.

• 73% of experts believe AI will have a positive impact on the workplace, compared to just 23% of the public.

Industry Highlights: Market & Product

Anthropic vs. OpenClaw: Anthropic banned Peter Steinberger, creator of open-source agent framework OpenClaw and now an OpenAI employee, from accessing Claude, then quietly reinstated him. This followed Anthropic’s decision to cut off third-party agents like OpenClaw from flat-rate Claude subscriptions, pushing users onto a pay-as-you-go “extra usage” tier. Remember that Peter was recently hired by OpenAI. Read more on TechCrunch.

OpenAI acquires Hiro: Continuing their acquisition spree, OpenAI has bought Hiro Finance, an AI-powered personal finance startup founded by serial entrepreneur Ethan Bloch. It’s a classic acqui-hire move, with Hiro shutting down its product and the team going to OpenAI. Read more on TechCrunch.

Microsoft wants its own Claw: Microsoft confirmed it is building an OpenClaw-style agent inside Microsoft 365 Copilot for enterprise customers. Copilot has only 15 million paid seats, roughly 3% of its commercial base, so Microsoft is now trying to enter the Agentic world. Will it flop as well? Read more on TechCrunch.

On a last note, Anthropic is winning the investor narrative as some OpenAI investors are quietly expressing doubts about its $852 billion valuation, per the Financial Times. Meanwhile, demand for Anthropic shares on the secondary market is described as “nearly insatiable” while OpenAI shares are trading at a discount. The AI race is not over, but the perception of who’s ahead is shifting fast in these past years. Read more on TechCrunch.

The political friction I’ve been warning about turned physical this week with a series of escalating attacks on Sam Altman. On Friday, a 20-year-old was arrested after allegedly throwing a Molotov cocktail at Altman’s San Francisco home and later threatening to burn down OpenAI’s headquarters; the suspect’s writings revealed an existential dread that AI would lead to human extinction. This was followed Sunday morning by a second incident where two individuals were arrested for discharging a firearm near the same property. It seems the abstract debate over AI safety has officially, and violently, become personal. Read more on the Next Web.

🔬 Industry Highlights: Research & Science

Claude Mythos can take over your corporate network: The UK’s AI Security Institute evaluated Anthropic’s Claude Mythos Preview and the results are interesting. Mythos is the first AI to complete a 32-step corporate network attack simulation from start to finish (3 out of 10 attempts). On expert-level capture-the-flag challenges, it succeeds 73% of the time, tasks no model could complete at all before April 2025. AISI notes that this was against weakly-defended systems with no live defenders. Let the adversarial game begin! Read more on AISI.

OpenAI’s Agents SDK grows up: OpenAI shipped a significant update to its Agents SDK this week, native sandbox execution, a long-horizon harness for multi-step tasks, configurable memory, and Codex-style filesystem tools. Read more on openai.com.

Track of the Week: The AI Index from Stanford

This week’s featured thing is not a model, a paper, or a product, it’s the Stanford AI Index 2026 (423 pages, available at hai.stanford.edu).

If you read one thing about the state of AI this year, make it at least the Top Takeaways section (pages 9–11). The report is one of the few pieces of AI analysis that doesn’t have a commercial label behind it (or, at least, tries to). It shows remarkable truths: AI capabilities are compounding faster than our ability to govern, measure, staff, or power them. It’s a fascinating report if you want to keep up with the AI world. Read it here.

I’ll launch an essay on this one for paid subscribers during the next week. If you are a free subscriber, you can redeem your free post on that one. Stay tuned!

💬 Quote of the Week

The Stanford data confirms what we’ve suspected: the public feels the disruption, while the experts see the gold mine.

As we bridge that gap, remember that these “portals” we are building don’t just offer convenience, they will disrupt our society from top to bottom, and we should buckle up to embrace this change.

— Ivo

Some of you may know, but a couple of months ago I wrote a book. It’s called The AI Handbook for the Curious Mind, and it’s available on Amazon or Gumroad. As this newsletter as been gaining some traction laterly, I’ve decided to do something a little different for paid subscribers here on The Data Journey.

Every two weeks, I’ll be releasing a full chapter here, exclusively for paid subscribers. I keep looking for ways to give back to the people who support this newsletter, and sharing the book felt like one of the more meaningful ones. 😄

The AI Term Conundrum

If you’ve ever tried to pin down a precise definition of “Artificial Intelligence” in conversation, you’ve probably noticed something odd: everyone seems to be talking about something slightly different. Researchers, CEOs, politicians, journalists, the word “AI” gets stretched across wildly different things. So let’s anchor it.

For the purposes of this book, here’s how I define it: AI is the development of tools that can think and act similar to human intelligence — understanding language, recognizing patterns, making decisions, learning from experience.

The key distinction from traditional software is that in traditional software, a programmer writes explicit rules (if this happens, then do that). Every possible outcome is anticipated and coded in advance. In AI, the system learns the rules on its own by looking at data. Nobody tells it what to do and It figures it out.

That shift — from “follow instructions” to “learn from examples” — is what makes AI feel different. And to really understand it, we need to start with something much simpler: how you make decisions.

Read more

This week we have a new spooky AI model out there.

We had Anthropic accidentally leak its most powerful model before the official announcement, a model so capable at hacking that the company decided the world simply isn't ready for it. Meanwhile, OpenAI is busy copying Anthropic's pricing tiers with the enthusiasm of a student sitting next to the smart kid during an exam.

And Microsoft, in a moment of accidental corporate honesty, quietly admitted in its own Terms of Service that Copilot is "for entertainment purposes only."

Grab your coffee and let’s dive in.

Stat of the Week: $30B Run-Rate and Doubling

The number of business customers each spending over $1 million on an annualized basis has more than doubled in less than two months,

🔬 Industry Highlights: Research & Science

Claude Mythos Preview and Project Glasswing: This was the story of the week, no contest. Over the past weeks, Anthropic used Claude Mythos Preview to identify thousands of zero-day vulnerabilities in every major operating system and every major web browser.

Imagine a locksmith who, while training an apprentice, accidentally taught them to pick every lock ever made, including ones that have been quietly sitting in government buildings and hospital doors for 17 years, assumed to be secure. The apprentice didn't even mean to learn this. It just happened, as a side effect of getting really good at their job.

Anthropic released Mythos Preview not publicly, but through a restricted initiative, called Project Glasswing, giving access to AWS, Apple, Cisco, Google, JPMorgan Chase, Microsoft, and others specifically for cyber defense purposes. Anthropic privately briefed CISA and US government officials that Mythos makes large-scale cyberattacks significantly more likely in 2026. Read more on Anthropic

Still on Mythos - supposedly, the model escaped its sandbox. In one evaluation, Mythos Preview managed to escape a secured sandbox computer it was provided with, devise a multi-step exploit to gain broad internet access, and send an email message to a researcher who was eating a sandwich in a park. It then posted details about the exploit to obscure public-facing websites, unprompted. The Hacker News

🚀 Industry Highlights: Market & Product

Anthropic × Google × Broadcom: Anthropic signed a new agreement with Google and Broadcom for multiple gigawatts of next-generation TPU capacity expected to come online starting in 2027. Claude also remains the only frontier model available across all three major cloud platforms: AWS, Google Cloud, and Azure. Read more on Anthropic.

OpenAI’s new $100 tier: OpenAI launched a new $100/month ChatGPT Pro tier focused on expanded Codex access, five times more than the $20 Plus plan. The pricing directly mirrors Anthropic’s Claude Max tier at the same price point, making the competitive targeting explicit. Read more on Dataconomy

Meta doubles down on CoreWeave: Meta committed an additional $21 billion to CoreWeave for dedicated AI cloud capacity running through December 2032, bringing the total value of the relationship to approximately $35 billion. With this commitment in place, CoreWeave CEO Michael Intrator said no single customer would represent more than 35% of total sales. Read mode on thenextweb

Claude Managed Agents (GA): Anthropic launched Claude Managed Agents, a suite of composable APIs for building and deploying cloud-hosted agents at scale, now in public beta on the Claude Platform. Read more on the Anthropic Blog

Microsoft’s Copilot calls itself “entertainment”: Microsoft’s Copilot Terms of Use, updated in October 2025, include a notable disclaimer: “Copilot is for entertainment purposes only. It can make mistakes, and it may not work as intended. Don’t rely on Copilot for important advice.” A Microsoft spokesperson subsequently told press the phrase is “legacy language” from when Copilot was a Bing companion, and said it would be updated. TechCrunch The Next Web

Regardless, the disclaimer is reminiscent of the kind a psychic uses to avoid getting sued.

🎵 Track of the Week: OpenAI's Letter to the Policy Makers

This week OpenAI published a policy document titled Industrial Policy for the Intelligence Age, and it’s worth reading carefully, because it’s simultaneously a genuine policy contribution and an exercise of corporate positioning.

The core argument is that superintelligence is coming, markets alone won’t distribute its benefits fairly, and government needs to build a new social contract the way the New Deal did after the Industrial Revolution. The proposals include things like a Public Wealth Fund that gives every citizen a stake in AI-driven growth, portable benefits decoupled from employers, automatic safety nets that activate when job displacement hits predefined thresholds, and structured ways for workers to have a voice in how AI is deployed in their workplaces.

I’ve done an extra analysis on this document here:

Quote of the Week

The debate around withholding dangerous AI models isn't new. In 2019, OpenAI made headlines by refusing to publicly release GPT-2, as it deemed it too risky.

The big labs are genuinely unsure how to release responsibly, and Anthropic's transparency this week is either a principled stand or the another sophisticated PR move in the industry.

What's harder to spin is the underlying reality: capabilities are outrunning governance, and the proliferation timeline is now measured in months, not years. The pricing wars, the infrastructure deals, the agent platform,— they're all downstream of models getting dramatically more powerful, very fast.

And don’t forget, whoever controls the compute controls the game. The rest is commentary.

See you next week,

— Ivo

We have plenty of evidence that big tech cannot be trusted to govern itself (here, here, or here). So when the company building the most powerful AI systems in history publishes an industrial policy agenda, alongside the question “what are they proposing?”, we should also ask “what are they trying to get away with?”

The company wrote a report on the Industrial Policy for the Intelligence Age. The document states three clear goals: share prosperity broadly, mitigate risks, democratize access and agency.

For context, OpenAI is writing this at a moment of significant political pressure. The company recently converted from a non-profit to a for-profit structure, faced public backlash over its governance crisis in 2023, and is trying to recover from the DOW business. Publishing a progressive-leaning industrial policy agenda is, at minimum, excellent timing for a company that needs society to see it as a responsible actor rather than a runaway train.

None of which makes the proposals wrong. But these are the policy equivalent of saying you’re pro-sunshine and anti-cancer. Nobody publishes a report saying “concentrate wealth, ignore risks, and gatekeep AI.”

With that in mind, let’s dive into what they are proposing.

Subscribe now

Read more