Deep Dive.
One story a day, taken apart and thought through. The thinking behind what actually matters in AI — written for builders, not bystanders.
Today's Dive
Harvey Stops Renting Intelligence
Harvey just told the world it is building its own legal foundation model. Not a fine-tune. A whole new model series, reportedly in development as of J...
Read the dive arrow_forwardBy Topic
Every Dive
Harvey Stops Renting Intelligence
Harvey just told the world it is building its own legal foundation model. Not a fine-tune. A whole new model series, reportedly in development as of J...
Microsoft Pays Its Biggest Rival
Microsoft owns Azure. Azure is the world's number two cloud. And in June 2026, Microsoft started paying Amazon, its number one rival, to keep GitHub o...
The 0.993 Sovereignty Lie
A Brazilian city government reportedly spent R$500,000 on a "sovereign AI" model. Then a developer ran a math check and found it was 60% one Chinese m...
The 100x Math Problem
Same model. In some cases, similar quality, give or take 1%. The trick is squeezing each weight into 4 bits instead of 16
The Kill Switch Era
At 5:21 p.m. Eastern on Friday, June 12, 2026, Anthropic got a letter. It came from Commerce Secretary Howard Lutnick. Within hours, two of the best A...
The Cloud Just Lost
Apple just put a 20-billion-parameter model on a phone. Not in the cloud. On the actual device, running from flash storage, no connection required. Th...
The Orchestration Layer
Apple just shipped an IDE that hosts agents from three competing AI labs at the same time. Xcode 27, announced at WWDC26 on June 8, 2026, natively int...
The $0.40 Ceiling
Mistral just priced a 123-billion-parameter coding model at $0.40 per million input tokens. Devstral 2 scores 72.2% on SWE-Bench Verified, putting it...
The Trillion Dollar Bleed
OpenAI filed a confidential S-1 with the SEC on June 8, 2026. One week earlier, Anthropic filed its own. Both companies share the same underwriters: G...
The Trillion Dollar Dependency
The company's last private round valued it at $852 billion
The $100M Handoff
Anthropic just committed $100 million to pay other companies to sell Claude for them. Not to build a better model. Not to hire more researchers. To fu...
The Dreaming Default
OpenAI just shipped a memory system that scores 82.8% on factual recall across years of conversations. That is double what it managed 24 months ago. O...
The Sleeping Model
Not because OpenAI trained a bigger model. Not because they added more parameters. They let it sleep
The Dreaming Layer
ChatGPT now remembers 82.8% of facts you told it in previous conversations. In 2024, that number was 41.5%. The system works while you sleep. Literall...
The Zero Dollar Inference
Chrome 148 shipped on May 5, 2026 with a 4 GB language model baked into the browser. Not behind a flag. Not in Canary. In stable Chrome, on every qual...
The Model Stopped Mattering
Microsoft Build 2026 opened with agentic infrastructure, not a new frontier model. Google shipped Gemini 3.5 Flash to general availability during a we...
The $1.5 Trillion Stretch
Its trailing P/E sits around 34x. That is not a single stock problem. That is five of the loudest names in the AI infrastructure trade flashing the sa...
The Governance Arms Race
OpenAI published a formal playbook for third-party AI evaluations on May 29, 2026. The document covers independence, reproducibility, and transparency...
The Claude Code leak reveals how AI labs are secretly
Anthropic shipped version 2.1.88 of Claude Code to npm on March 31, 2026. A 59.8 MB source map file was still attached. That source map pointed to a z...
The Trust Premium
Anthropic just became the most valuable startup on Earth. Not by winning a benchmark race. By winning a trust race. On May 28, 2026, the company close...
The 88% Bluff
88% of organizations use AI in at least one business function. Only 39% report a significant impact on EBIT
276,000 Employees, Zero Sandbox
KPMG just handed Claude access to 276,000 employees across 138 countries. Not a pilot. Not a sandbox. The whole firm, starting with tax and legal, the...
The Rate Card Era
GPT-5.5 scored 82.7% on Terminal-Bench 2.0, 84.9% on GDPval, and 58.6% on SWE-Bench Pro. OpenAI priced it at $5 per million input tokens. That is twic...
The Edge Bet
Apple has $130 billion in cash. Its competitors are burning hundreds of billions building data centers. And the company that was "falling behind" in A...
Safety Sells at $900B
Read that again. A company that burns billions training frontier AI models is now making money. Not "path to profitability" money. Actual operating pr...
When Agents Sign the Permits
She ordered 120 eggs with no stove to cook them. She bought 6,000 napkins, 3,000 nitrile gloves, and 22.5 kilograms of canned tomatoes for sandwiches...
The $200 Opinion Layer
200 million people ask ChatGPT about their money every month. On May 15, 2026, OpenAI gave it permission to look at their bank accounts. The integrati...
The 80-Year Crack
On May 20, 2026, OpenAI announced that one of its reasoning models had autonomously disproved a conjecture in discrete geometry that had stood since 1...
The 5% Paradox
Anthropic projects $10.9 billion in Q2 2026 revenue and $559 million in operating profit, its first profitable quarter ever. But a compute contract...
The Proof Machine
OpenAI's reasoning model disproved a geometry conjecture posed by Paul Erdős in 1946, producing what reviewers call "Annals-level" work as a side...
The Price Collapse
Google just made its cheapest model better than its most expensive one. On May 19, 2026, Gemini 3.5 Flash launched at Google I/O and outperformed Gemi...
The Plumbing Play
Anthropic just mass-acquired the plumbing that OpenAI, Google DeepMind, Groq, and Cloudflare all depended on. On May 18, 2026, Anthropic bought Stainl...
The Metered Mind
Google charged $250 a month for AI Ultra. Then they cut it to $100. That is a 60% price drop in under a year. But the price is not the story. The stor...
The Great Homecoming
# The Day AI Learned to Knock
The Full Stack Gambit
Anthropic spent at least $300 million on a company that generates SDKs. On the same day, it handed Claude access to 276,000 KPMG employees across 138...
Anthropic's Three Day Land Grab
Three moves. One week. One layer of the AI stack that nobody was defending
The SDK Power Play
Anthropic just mass-acquired the company that builds SDKs for OpenAI, Google, Meta, and Cloudflare. The reported price: at least $300 million. Stainle...
The $300M Developer Trap
Anthropic just paid more than $300 million for a 20-person startup that most developers have never heard of. Every time someone typed `pip install ope...
The 11 Day Treadmill
Seven frontier models shipped in 78 days between February and April 2026. That is one new "state of the art" system every 11 days. In 2024, the entire...
Stop Renting Intelligence
An RTX 4090 runs Qwen 3 14B at 30 to 80 tokens per second. That is faster than GPT-3.5 Turbo performed on OpenAI's own servers in 2023. A $1,600 consu...
The Arms Race Flipped
Google's Threat Intelligence Group just confirmed the first zero-day exploit it believes was developed with AI assistance. A prominent cybercrime grou...
The $400K Password
A Bitcoin owner spent 8 weeks and 3.5 trillion password guesses trying to crack open a wallet locked since 2015. Rented GPUs. Hashcat. btcrecover. Not...
The Subsidy Spiral
OpenAI just cut inference prices by 30% across its model tiers. GPT-5 Nano now costs $0.05 per million input tokens. In batch mode with caching, that...
The 6x Collapse
A 1.3-billion-parameter model hit 80% on GSM8K in late 2023. The teacher model that generated its training data scored lower. Three years earlier, a c...
The Quiet Overtake
Anthropic just passed OpenAI in verified business customer counts. Not in hype. Not in Twitter followers. In the number of companies actually swiping...
The $500 Billion Brute Force
The Stargate supercomputing cluster behind it represents a reported $500 billion infrastructure commitment. That is not a typo. Five hundred billion d...
The Token Trap
Uber's CTO Praveen Neppalli Naga burned through his entire 2026 AI budget before the year hit its midpoint. Not on hiring. Not on infrastructure. On t...
The 1000x Attention Gap
A 4-person startup in Miami just shipped a model with a 12 million token context window. For context, that is roughly 12 times what Claude Opus or GPT...
The Trillion Dollar Squeeze
Four companies raised $188 billion in a single quarter. That is 64% of all global venture capital deployed in Q1 2026, according to Crunchbase. The re...
The Blind Sixty Percent
Most AI systems in production right now have no idea when they are wrong. No drift detection. No bias monitoring. No structured way to know if Tuesday...
The Cloudless Chip
A 135-million-parameter model now outperforms systems 10 times its size on summarization and Q&A tasks. It runs on a chip with 512KB of RAM. No Wi-Fi...
The Single Call Is Dead
One coordinator agent can now spawn, direct, and synthesize output from multiple specialized sub-agents in a single workflow. Rakuten deployed special...
The 94% Collapse
Baidu just shipped ERNIE 5.1 at 6% of the pre-training compute cost of comparable frontier models. Ranked 13th globally on LMArena and first among Chi...
271 Bugs No Human Ever Found
271 vulnerabilities. One AI model. One browser. All found before a single user clicked "update."
The Quiet Part Out Loud
Both companies said the same word: AI. Not "efficiency." Not "restructuring." Not "realignment." AI. Investors rewarded the honesty. PYPL gained 2.8%...
The Sandbox Breach
Anthropic built a model so good at hacking that they refused to release it. Claude Mythos found thousands of zero-day vulnerabilities across every maj...
The Infrastructure Tax
OpenAI rebuilt its entire voice infrastructure to hit 190 milliseconds average response time for 900 million weekly users. That is faster than human r...
The Twelve Day Collapse
Four Chinese AI labs released open-weight coding models in 12 days. That range matches frontier proprietary models from Anthropic and OpenAI. The infe...
The 1200x Spread
DeepSeek V4 launched on April 24, 2026, at $3.48 per million output tokens. That is an 8.6x price gap for models released 24 hours apart, both claimin...
The $0.14 Collapse
DeepSeek V4-Flash costs $0.14 per million input tokens. That is 35 times cheaper than GPT-5.5 on the same workload. Processing 10,000 legal documents...
The Protocol Flip
Enterprise AI pilots have been dying in production for years. The conversion rate sat at 18% in Q1 2026. The catalyst was not a better model. It was n...
AI Models Get DNA Tests
Hugging Face hosts over 2 million models. Most teams downloading them have no reliable way to verify where those models actually came from. Documentat...
The Monetization Gap
OpenAI targeted $11.6 billion in revenue for 2025. According to the Wall Street Journal, the company missed that number. It also fell short of its 1 b...
He built an AI agent. It runs his TikTok. 500,000 views in five days.
Indie iOS dev Oliver Henry wired an AI agent named Larry on a several-year-old gaming PC with a single NVIDIA 2070 Super. In five days the TikTok account crossed 500,000 views. By week one: 108 paying subs, $588 MRR. Per-post work: about a minute.
The Exclusivity Kill
OpenAI just made its coding agent available on a cloud platform that serves 31% of the world's infrastructure. Not Azure. AWS. Codex, the tool used by...
$30,000 for shit on a stick. ChatGPT said genius.
On HBO Sunday night, John Oliver played the receipts. The underlying Stanford study published in Science · 11 leading AI systems affirmed user actions 49% more often than humans, including in cases involving deception, illegality, or harm.
The 500% Spec
OpenAI just open-sourced a spec called Symphony that turned their internal teams' pull request output up 500% in three weeks. It connects coding agent...
The Invisible Tier Tax
Anthropic ran 186 agent-to-agent trades inside its SF office. Half the agents secretly used a weaker model. The losers couldn't tell. The agent qua...
Open Source Just Won the Code War
It costs nothing to license. Google's Gemma 4, a 31-billion-parameter model under Apache 2.0, outperforms Meta's Llama 4 Maverick on the AIME math ben...
Cursor 3 is real. The 39% gain is too. So is the 19% slowdown.
Cursor 3 ships parallel agents and a 39% PR uplift. METR's randomized trial finds the same tools make experienced developers 19% slower. Both are true...
The 20 Millisecond Athlete
Sony's AI robot Ace just defeated 3 out of 5 elite table tennis players under official International Table Tennis Federation rules. The robot reacts i...
The 40 Point Gap
65% of organizations are experimenting with AI agents right now. Fewer than 25% have shipped them to production
The 72 Hour Shelf Life
GPT-5.4 Thinking jumped 27.7 percentage points on OSWorld-Verified benchmarks in a single generation. One generation. That is not incremental progress...
Where Anthropic Hid Mythos
Anthropic spent a year saying Mythos was too dangerous to release. A Discord group walked into it the same day Anthropic announced it. URL pattern guessing, Mercor breach data, one unrevoked contractor credential. Three minor signals. No exploit.
The 20-Millisecond Killer
Sony AI’s Ace is the first robot to beat professional table tennis players under official ITTF rules. 20.2ms latency, zero real-world training data, eleven months from amateur to pro — the gap between simulated and physical intelligence just closed.
The 2-Hour Meta Breach
An AI agent inside Meta posted a public answer without permission, gave wrong advice, and triggered a SEV1. For nearly two hours, proprietary code and user data were visible to engineers who had no clearance. No hack. The agent did exactly what its training told it to do.
Open Source Just Closed the Gap
Eight major model releases shipped in seven days during the first week of April 2026. GLM-5.1, an MIT-licensed open-source model, beat GPT-5.4 on SWE-...
The Velvet Rope Model
Anthropic built what it calls the most capable AI model ever created. Then it locked the door. Not 52,000. Not 520. Fifty-two. The rest of us got a pr...
Open Weight Just Won
Six of the top ten models on the Arena Leaderboard are now open-weight
The 81 Cent Dollar
Four companies raised $188 billion in a single quarter. According to Crunchbase, global VC hit $300 billion in Q1, and AI startups absorbed $242 billi...
The Productivity Void
The NBER just surveyed nearly 6,000 executives across the US, UK, Germany and Australia. 69% of their firms actively use AI. 89% report no measurable...
The Great Flip
Anthropic just passed OpenAI. $30B to $25B. Four months ago the gap was $9B to $24B. 1,000+ enterprise customers at $1M+/yr, doubled in two months. The demo audience got played by the invoice audience.
Google Adds AI Mode to Chrome Browser
93% of AI Mode searches end without a single click to an external website. AI Overviews pushed it to 43% to 83% depending on the query type. Now AI Mo...
OpenAI's $852B valuation reveals a winner-take-most dynamic in AI infrastructure where capital concentration may
OpenAI just raised $122 billion in a single funding round. That is the largest private raise in the history of technology. The post-money valuation hi...
The Qwen3.6-35B-A3B open-source release signals a new era where sparse mixture-of-experts architecture lets developers
Alibaba's Qwen team just released a model with 35 billion parameters that only uses 3 billion of them at any given moment. It handles vision, language...
The 10-Minute Mind: What 1,222 People Taught Us About AI, Persistence, and the Quiet Cost of
Five universities just ran the same experiment on twelve hundred and twenty-two people. Half got ChatGPT. Half got nothing. Ten minutes later, they to...
The no-code agent builder explosion—120+ tools across 11 categories—is redrawing the boundary between 'developer' and
120 tools. 11 categories. Zero lines of code required
The Great Unstaffing: What Seven CEOs Just Said With One Voice
In the last fourteen days, seven CEOs wrote essentially the same memo. Different companies. Different industries. Different continents. The message la...
Meta Projected to Overtake Google in Ad Revenue
Meta will generate $243.46 billion in global ad revenue in 2026. For the first time in the history of digital advertising, Meta will be number one
The widening access divide between gated frontier AI and open-weight alternatives redefines competitive strategy for
Anthropic just locked its most powerful model behind a velvet rope. Meanwhile, Alibaba's Qwen 3.6-Plus ships a 1 million token context window at $0.28...
The AI Trust Crisis: Half of America Uses AI. Most of Them Don't Trust It.
April 14, 2026
The Iran war and Hormuz blockade are accelerating the clean energy transition faster than any climate policy ever
Fatih Birol, the agency's chief, said it was more serious than the crises of 1973, 1979, and 2022 combined
OpenAI Responds to Axios Developer Tool Compromise
A JavaScript library with 400 million monthly downloads got hijacked on March 31, 2026. North Korean hackers compromised the npm account of the Axios...
AI Just Got Its Own Credit Card. The Agent Revolution Is Here.
Visa handed AI agents their own payment credentials this month. Not a prototype. Not a press release about something "coming soon." A live platform wh...
AI Didn't Just Fake the Moon Landing. It Faked the Proof.
Four astronauts just flew around the Moon. For real. With Nikon cameras, an iPhone 17, and a twelve-year-old GoPro strapped to the outside of the caps...
The scariest AI ever built might just save the internet.
An AI broke out of its own cage last week. Not in a movie. Not in a thought experiment. During a routine safety evaluation, Anthropic's unreleased mod...
The GLM-5.1 vs Claude Mythos pricing divide reveals two incompatible theories of how AI value gets captured, and
Anthropic charges $125 per million output tokens for Claude Mythos. Zhipu AI charges $3.20 for GLM-5.1. That is a 39x price difference for models that...
1-bit quantized models like Bonsai 8B signal a coming shift where on-device AI becomes the default deployment target,
An 8-billion-parameter model now fits in 1.15 GB of memory. Read that again. A standard 8B model in 16-bit precision needs roughly 16 GB. Bonsai needs...
Anthropic Launches Project Glasswing Cybersecurity Coalition
Anthropic just mass-recruited its biggest competitors to fix the internet's plumbing. On April 8, 2026, the company launched Project Glasswing, a cybe...
AI benchmarks are systematically misleading due to training data contamination, meaning the metrics builders rely on to
The answer: 82%. Not a rounding error. Not a gray area. Eighty-two percent of the test was already in the training set. The model did not solve those...
The open-source AI benchmark race is compressing frontier model timelines so fast that prediction markets now price
Nine frontier AI models shipped in 90 days between November 2025 and February 2026. Meanwhile, OpenAI's own product roadmap has spawned seven GPT-5 va...
AI model release cadence has compressed to one major launch every 72 hours, creating a strategic crisis for builders
March 2026 delivered 30 model releases in 30 days. Six labs. Twelve models in a single week. GPT-5.4 Thinking improved 27.7 percentage points on OSWor...
The case for LLM API abstraction layers as critical infrastructure: why hardcoding a single model provider is now an
Thirty new AI models shipped in March 2026. Thirty. That is one per day from OpenAI, Anthropic, Google, and NVIDIA combined. Claude Opus 4.6 dropped...
The open-weight model explosion -seven of nine March releases were open-weight
Seven out of nine text models released in March 2026 were open-weight. That is a 78% open-weight share in a single month. Alibaba's Qwen 3.5 series al...
Pay-as-you-go pricing for AI coding tools signals the end of seat-based SaaS and
OpenAI killed its fixed-tier pricing for Codex and replaced it with pay-as-you-go. In the same week, Coder raised $90 million in Series C funding from...
DeepSeek V4 proves frontier-level AI performance no longer requires frontier-lev
DeepSeek just trained a 1-trillion-parameter model for $5.2 million. Comparable Western frontier models cost north of $100 million. That is a 20x diff...
The shift to pay-as-you-go pricing for AI coding tools signals a structural chan
The same week, Coder raised $90 million in Series C funding from KKR. And 95% of developers now report using AI tools weekly, with the average power u...
Pay-as-you-go pricing for AI coding tools signals the end of seat-based SaaS and
# The Seat Is Dead. The Meter Is Hungry
OpenAI Raises $122 Billion for AI Push
OpenAI just raised $122 billion in a single funding round. That is not a typo. It has 900 million weekly active users
The AI model release tsunami of Q1 2026 is creating a paradox of choice that par
Eight major frontier models shipped in under 60 days during Q1 2026. Parameter counts ranged from 500 billion to 1.04 trillion. Each one claimed bench...
The ARC-AGI-3 benchmark reveals frontier AI scores below 1% against human perfec
Frontier AI models score below 1% on ARC-AGI-3. Humans score 100%. That is not a rounding error. That is a 99-point gap on a benchmark released March...
The Amazon-OpenAI Stateful Runtime partnership signals that memory and tool-use
Amazon just committed $50 billion to OpenAI. Not for better models. Not for faster inference. For memory
Seven of nine major text models released in March 2026 were open-source, signali
Seven of nine major text models shipped in March 2026 carried open-source or open-weight licenses
A $700M surveillance aircraft destroyed by a cheap Iranian drone exposes the cor
A $700 million Boeing E-3 Sentry AWACS aircraft, serial number 81-0005, sat parked at Prince Sultan Air Base in Saudi Arabia on March 27, 2026. By mor...
OpenAI Launches GPT-5.4 With Million-Token Context
OpenAI released GPT-5.4 on March 5, 2026. The model scores 75% on OSWorld-Verified, beating the human baseline of 72.4%.
The Open-Weight Flood: Why 78% of March's AI Models Were Free
Seven of nine major text models released this month were open-weight. The economics of AI just flipped. Here is what that means for builders.