Blog

On 27 January 2025, Nvidia lost roughly $589 billion in market value in a single trading session. The trigger was DeepSeek-R1, a reasoning model from a Chinese lab that matched far more expensive systems at a fraction of the price. Two days later, OpenAI said it had seen evidence that DeepSeek had used its models to train its own, through a technique called distillation. Most people had never heard the word. Within weeks, it was a question for markets, regulators, and courts.

Inspiration On 27 January 2025, DeepSeek-R1 wiped $589B off Nvidia's market capitalization in a single session — the largest one-day loss in stock-market history. Two days later, OpenAI accused DeepSeek of distilling its models. Within weeks, distillation became an export-control question, a Senate-hearing question, and an EU AI Act question. Distillation is the quiet engine behind most models people actually pay for — GPT-4o-mini, Gemini Flash, Claude Haiku, Llama 3.2, the e

Most people who use an AI assistant every day have never heard the word "distillation," and yet it shapes nearly everything they pay for. The fast, affordable models that answer customer emails, draft code, and summarize documents are rarely the giant frontier systems that make headlines. They are smaller students, trained to imitate a much larger teacher. That handoff of capability from big to small is knowledge distillation, and until recently there was no single place to u

What a 0.6-billion-parameter student learns in 20 minutes, and what an AI agent found out about it overnight. Code, logs, and results: https://github.com/alessoh/ai-distillation Summary We took the distillation recipe from chapter 8 of Sebastian Raschka's *Build a Reasoning Model (From Scratch)*, wrapped it in Andrej Karpathy's AutoResearch loop, and let a coding agent run experiments on an H100 for a night. The student is Qwen3-0.6B. The teachers are DeepSeek-R1 and Qwen3-2

Connect Four was one experiment on the research loop. Here are ten more — each chosen to break a different part of. game metric status stresses nim win_rate queued greedy selection vs grokking kuhn_poker exploitability queued non-transitive strength 2048 mean_score queued noisy fitness signal ultimate_ttt win_rate queued hierarchy in the blueprint hex_7x7 win_rate queued inherited inductive bias othello_6x6 win_rate queued internal world models gomoku_9x9 win_rate queued tran

Here is an experiment you can run tonight. Give an AI agent a Connect Four training script, a fixed evaluation function, and one instruction document. Tell it to modify the script, train a network for five minutes of self-play, check whether the agent plays better, keep or discard the change, and repeat. Then go to sleep. In the morning there is a ledger — results.tsv, one row per experiment, roughly a hundred rows — and a game-playing network measurably stronger than the one

For decades, recursive self-improvement — AI that makes itself smarter, which makes it better at making itself smarter — lived in philosophy papers and forecasting debates. In the summer of 2026, it became a crowded empirical subfield. Across July and August alone, roughly eighty arXiv preprints attacked the problem directly, two independent surveys each organized about 1,250 papers on the topic, and — most tellingly — the field began publishing rigorous critiques of its own

If you have used an AI coding assistant, a research agent, or any tool that lets a model browse the web, run code, and finish a multi-step job on its own, you have already benefited from something the industry calls scaffolding. The model itself is only half the story. The other half is the orchestration layer wrapped around it, and in 2026 that layer has become the primary battleground of applied AI. What Scaffolding Actually Is A large language model, on its own, is a text-

In the next decade, exponential technology will collide with entrenched power. The next step is collision as a management problem. For each critical choke point, the questions are practical ones. What does failure actually look like? What are the early warning signs? And what are the best available strategies for handling it, whether you are a policymaker trying to keep a society stable, a builder trying to keep the technology moving, or a citizen trying to keep the benefits

There is a story technologists love to tell about the next ten years, and it goes something like this. Artificial intelligence keeps doubling in capability on a cadence measured in months, not decades. Robots step out of the factory cage and into warehouses, kitchens, and hospitals. Self-driving trucks quietly take over the interstates at three in the morning, and robotaxis make car ownership feel as quaint as owning a horse. Reusable rockets drop the cost of reaching orbit u

On July 21, 2026, Jack Dorsey's Block released Buzz, a free, open-source collaboration platform that looks, at first glance, like yet another Slack clone. It has channels, threads, direct messages, voice huddles, and media sharing. Anyone who has lived inside a modern team chat tool will feel instantly at home. But the familiar surface conceals a genuinely different architecture, and that architecture is the story. Buzz was not built for teams of people who occasionally summo

Every threat in an intelligence feed carries a severity score, and most of the time the scores blur together. Then, occasionally, one arrives that earns every point it is given. CVE-2025-6514 is that kind of threat. It sits at the top of the ThreatPulse feed with a CVSS score of 9.6, and it deserves the ranking because it turns one of the most ordinary actions in the agent world, connecting to a remote tool server, into a complete takeover of the machine doing the connecting.

For thirty years, cybersecurity has spoken a shared language. We classify a threat as ransomware or phishing, a zero-day or a supply-chain compromise, and everyone in the room knows roughly what is meant, what it targets, and what to do about it. That shared vocabulary is not a convenience. It is the machinery that lets defenders triage, prioritize, and respond at speed. You cannot defend what you cannot name, and for a generation we have had good names. Autonomous AI agents

When people picture AI agents working together, they often imagine a quiet exchange of private messages, one agent asking another for help and waiting for a reply. That picture is real, but it is only half the story. Increasingly, agents do not just message one another one-on-one. They gather in shared conversations that look a great deal like the group chats on your phone, and the difference matters more than it first appears. The most common way agents coordinate today is t

ThreatPulse is live at threatpulse.dev — a daily, read on how agents are being attacked. This site keeps returning to one question: not what AI agents can do, but what they should be allowed to do — and who is watching when they do it. Today we're adding something to that second part. We built the watch. ThreatPulse, now live at threatpulse.dev, tracks the security threats specific to the agent era. Not the abstract worry that "AI might be dangerous," but the concrete, alread

A young protocol for agents to talk to each other has crossed an important line. That is worth marking, and worth keeping in proportion.

A payment idea the web set aside decades ago has quietly come back to let software buy from software. Here is what it really changes, and what it does not.

It is tempting to treat identity as a detail to sort out later. That instinct is exactly backward, and here is why everything else depends on it.

AI is the fastest-adopted technology in human history. ChatGPT hit 900 million weekly users in under three years. Nearly 40% of American adults now use generative AI regularly, a penetration rate the internet took half a decade to achieve and the personal computer needed more than a decade to match. And yet, Goldman Sachs reported in March 2026 that AI has contributed essentially nothing to economy-wide productivity growth. The number they put on it was 0.1 to 0.2 percentage

You launch an AI agent on a Tuesday evening and go to bed. While you sleep, the agent runs eighty experiments on the code that built it, keeps the improvements, discards the failures, and starts again. By Wednesday morning, it has found optimizations you missed for months. You open the log and read the night's work. Somewhere around experiment sixty, the agent stopped tweaking and started redesigning. The architecture it produced is better than yours. It is also better than t