A weekly briefing on AI · Markets · Power · by Kaloian Parchev
Week 24 June 5–11, 2026
The car has no brake pedal
AI wrote 80% of Anthropic’s own code. The lab that built the accelerator is now asking for a way to stop.
S&P 500
7,267
−1.62%
Jun 10 close · Yahoo Finance
Nasdaq
25,169
−1.98%
Jun 10 close · Yahoo Finance
Bitcoin
$61,672
−2.3%
Jun 10 open · Yahoo Finance
Gold
$4,104
−0.70%
Jun 10 · Yahoo Finance
This week the market fears:
AI-driven deflation meeting Middle East energy price spikes
Est. reading time: 7 min
The week in 60 seconds
Anthropic told the world its AI already writes 80% of its own code — and called for a global brake mechanism before the next step, recursive self-improvement, makes that irreversible. Meanwhile, AI CEOs from OpenAI, Anthropic, and Microsoft set rivalry aside to warn Congress that AI is now lowering the barrier to bioweapons. Google quietly laid off cybersecurity researchers at Mandiant and its Threat Intelligence Group, redirecting capital to AI. And Anthropic published two science papers showing Claude matching specialized chemistry software on molecular analysis and exposing fragile data infrastructure as the real bottleneck for AI in biology.
Big Idea
The accelerator is asking for a brake
Anthropic co-founder Jack Clark said publicly what most frontier labs have quietly known for a while: the AI industry has a gas pedal but no brake pedal. The company’s new research paper, published through its Anthropic Institute, maps the path from AI coding assistant to AI that builds itself — and argues that a verifiable global pause mechanism should exist before that threshold is crossed.
The evidence isn’t hypothetical. As of May 2026, more than 80% of code merged into Anthropic’s own production codebase was written by Claude. Engineers were merging 8x more code per day in Q2 2026 compared to 2024. That’s not AI as a tool. That’s AI as a co-developer — and accelerating.
The company that built the fastest car is now saying we need a working brake before someone else floors it.
What makes this week’s announcement structurally different from the 2023 pause letters is the source. The Future of Life Institute letter was signed by researchers and outsiders. This one comes from a lab with firsthand knowledge of current capability levels — and from leadership who openly acknowledge that a unilateral pause would be commercially suicidal unless it’s coordinated. Clark’s proposed solution: internationally verifiable mechanisms that would allow multiple labs to halt simultaneously, modeled loosely on Cold War arms control logic.
The honest read: this is both a genuine safety argument and a coordination problem Anthropic can’t solve alone. But the fact that the accelerator is asking for a brake changes the political landscape for AI governance more than any external advocacy campaign has managed to date.
The Model
THE RECURSIVE TRAP
1. Lab builds AI to write code faster → engineers become 8x more productive
↓
2. AI writes 80% of the lab’s own production code → AI improves AI
↓
3. Next step: AI builds and trains successor models autonomously
↓
4. Alignment bugs in the parent model compound into each successor
↓
5. Humans are pushed to the margin of a process they used to control
The dangerous moment isn’t when AI becomes smarter than humans — it’s when no single human or institution can verify it anymore.
By the Numbers
AI-WRITTEN CODE SHARE AT ANTHROPIC — ESTIMATED TRAJECTORY
% of production code written by Claude · 2024–2026 · Anthropic estimates
Source: Anthropic / The Next Web (thenextweb.com) · Intermediate values estimated from stated trend; Q2’26 = confirmed 80%
The curve isn’t linear — it’s accelerating. Going from 0% to 30% took roughly 18 months; going from 30% to 80% took roughly the same. If the pattern holds, reaching 100% isn’t a distant scenario — it’s a planning horizon.
NMR PREDICTION ACCURACY — MEAN ABSOLUTE ERROR (¹H, ppm)
Lower is better · Hydrogen shift prediction across 20 compounds · Anthropic white paper, Jun 2026
Source: Anthropic, “Making Claude a chemist” white paper (anthropic.com/research/making-claude-a-chemist) · Jun 5, 2026 · Intermediate values are approximations from published figures
Opus 4.7 beats dedicated chemistry software — ChemDraw and MestReNova — on hydrogen NMR prediction accuracy. The tolerance window is ±0.20 ppm; Opus 4.7 lands at ±0.079, less than half. What matters is not the benchmark score but the implication: general-purpose AI is now outperforming specialized scientific software in its own domain.
Using the DRC Ebola outbreak as a case study, Anthropic shows that fragile database infrastructure — not model capability — is the real bottleneck for AI in science.
The same week the stat dropped, Anthropic published a paper arguing the world needs a verifiable off-switch before recursive self-improvement becomes irreversible.
Google cut elite human security researchers the same week it unveiled new AI-powered security tools — the clearest single-week illustration of the AI substitution thesis in action.
Crypto Pulse
Bitcoin
BTC fell from ~$73,500 at the start of June to briefly dip below $60,000 over the weekend — its lowest level since 2024 — before stabilizing around $61,700–$63,000 by mid-week. The slide was driven by geopolitical energy concerns from the Middle East, capital rotating into AI equities, and a large-holder sell-off that triggered cascading liquidations.
Ethereum
ETH tracked Bitcoin’s decline, dropping from $2,000 at the start of the month to roughly $1,638 by June 10. The ETH/BTC ratio held roughly steady, suggesting no internal rotation between the two — both simply fell together on macro risk-off sentiment.
The macro read: analysts are noting crypto may offer diversification from AI-dominated equity flows — but this week demonstrated that when macro risk comes from the same direction (energy prices, rate expectations), crypto and equities both get hit simultaneously, regardless of their different narratives. Source: Yahoo Finance, CoinDesk.
What I’m Reading
The Coming Wave
Mustafa Suleyman · 2023 · Crown Currency
The recursive self-improvement debate brought me back to this — Suleyman’s argument that the central challenge of our era is the “containment” of powerful technologies is exactly the lens Anthropic is now using publicly. What I find correct is his framing that containment is not a binary switch but a continuous negotiation between capability and oversight. What I find too neat is his confidence that nation-states can coordinate fast enough to matter — this week’s congressional letter shows how fragile even tech-company alignment is. What I’m taking from it: the window for building brake pedals isn’t infinite, and the people who waited for AI to “get close enough” to act are already behind.
One Number
8×
Engineers at Anthropic were merging 8 times more code per day in Q2 2026 versus Q2 2024 — driven entirely by AI-generated output. This isn’t a productivity boost. It’s a structural change in what software engineering means at a frontier lab. Source: Anthropic / The Next Web, June 2026.
This Week I Noticed
Something I’ve been sitting with: when a company as commercially motivated as Anthropic publicly asks for a way to slow itself down, the honest reaction isn’t “how reassuring” — it’s “what have they seen internally that makes them willing to say this out loud?” Safety arguments that cost nothing are cheap. This one isn’t cheap. If I’m reading the week correctly, the brake pedal conversation is moving from philosophy to engineering very quickly.
On the science side: Anthropic’s biology paper used the DRC Ebola outbreak — more than 200 deaths confirmed as of late May — as a case study for why AI agents can’t reliably retrieve viral sequence data from NCBI without specialized retrieval layers. The point wasn’t to be grim; it was specific: when infrastructure designed for human clicks meets AI agents that need APIs, researchers lose days that don’t exist in an outbreak. The bottleneck to AI-accelerated science isn’t model intelligence. It’s that the databases science runs on were built before AI was a user. Source: Anthropic Research, June 8, 2026.
Quote
“The AI industry right now has a gas pedal, but it doesn’t have a brake pedal in the car, and we want to do some of the work to build that pedal.”
— Jack Clark, Co-founder, Anthropic · BBC interview, June 5, 2026
The reason this quote lands harder than most AI safety statements: Clark is describing his own company’s car. He’s not a regulator or an academic. He’s the one who helped build it — and he’s saying it publicly while the engine is running.
Three things to remember
The AI-writes-its-own-code milestone is no longer theoretical — 80% at Anthropic as of May 2026 — and the political window to build oversight infrastructure is narrowing faster than most governance timelines assume.
General-purpose AI models are now outperforming specialized scientific software in their own domains, which changes the economics of every sector that has historically been protected by technical complexity.
Google’s decision to cut human security researchers while launching AI-powered security tools is the clearest signal yet that the AI substitution thesis is moving from white-collar knowledge work into elite, technically specialized roles.
One Thing To Do
Look at any company in your portfolio with a specialized data moat — legal databases, financial data, scientific repositories. Ask whether their competitive advantage depends on humans needing to click through their interface, or whether it survives when AI agents become the primary users. The biology paper makes clear: that distinction will separate durable businesses from ones quietly being disintermediated.
Support This Newsletter
Sponsor
Trading 212
No account yet? Use my link — you get a shot at free shares, and you help keep this newsletter going. No catch.
