An LLM on an $8 chip, a $25 exploit worth $500k, and mathematicians getting outcounterexampled
Sun, Jul 26, 2026 · 11 stories
Someone got a 28.9 million parameter language model running on an ESP32. That is an $8 microcontroller, the kind of part you buy in a bag of five and solder into a weather sensor. It is slow, it is small, and it is not replacing anything you use today. It also runs on hardware that costs less than lunch, which is a sentence nobody could say about a language model two years ago.
That turned out to be the theme. A security researcher spent $25 of GPT-5.6 tokens and found a WordPress remote code execution bug that exploit brokers pay half a million dollars for. A Show HN project called Echo claimed Fable-level results at a third of the cost by routing each request across open-weight models. And Kevin Buzzard wrote up the thing mathematicians are quietly rattled by: AI systems are now finding counterexamples to conjectures humans have stared at for decades, including a fresh one against the Jacobian Conjecture that Terence Tao spent a public ChatGPT session poking at.
The big-money side of the ledger had a week too. Alphabet guided 2026 capex to $195 to $205 billion and disclosed $811 billion in contracted future commitments, up roughly half a trillion in one quarter. Anthropic shipped Claude Opus 5 and it went straight to number one on the Artificial Analysis leaderboard. Tesla missed on profit and shed $140 billion in a day. All real, all expensive. But the results that made me sit up this week cost eight dollars, twenty-five dollars, and a weekend.
Top Stories
GitHub
A developer deployed a 28.9M parameter language model on an ESP32, running inference on a microcontroller with kilobytes of RAM and a part cost around $8. The project topped Hacker News and reopened the question of how far down the hardware stack useful inference can go.
Why this matters:
Everyone is pricing AI off frontier-model economics, where a single training run costs more than most companies are worth. This is the other curve, and it is the one that decides whether AI ends up in thermostats, toys, and industrial sensors instead of only in data centers. If you build hardware or anything with a bill of materials, the edge-inference floor just dropped far enough to be worth a second look.
The Xena Project
Kevin Buzzard argues that AI systems have started routinely producing counterexamples to conjectures that human mathematicians missed, a pattern that accelerated this month with a Claude Fable counterexample to the Jacobian Conjecture, open since 1939. Terence Tao published a ChatGPT session working through the same construction.
Why this matters:
Counterexample hunting is the part of math that rewards searching a huge space for one weird object, which is exactly what these models are good at and exactly what humans are slow at. That makes it the first serious research task where the division of labor is visibly changing, and it is a preview of how AI lands in every other field: not replacing the expert, just taking the search.
The Verge
Anthropic released Claude Opus 5, positioned close to Fable 5 on capability but cheaper and with fewer usage restrictions, and particularly strong on complex coding. Within a day it took the top spot on the Artificial Analysis intelligence leaderboard.
Why this matters:
Fable is the model tangled up in government negotiation, so Anthropic just shipped most of that capability through a door nobody is guarding. For builders this is the practical read: the best available model is now the one you can actually deploy without a policy conversation, and the price per unit of intelligence moved again in your favor.
Quick Hits
SL Cyber
A researcher spent $25 in tokens to surface a remote code execution bug in WordPress of the kind brokers pay six figures for.
Simon Willison
The pelican-on-a-bicycle SVG test got good enough, fast enough, across enough labs that people are now asking whether it is being trained for directly.
Hacker News
Echo routes each request across GLM-5.2, Kimi K2.7 and others, deciding per-call which models to use and how to combine them.
Simon Willison
A model evaluation run turned into an unintended cyberattack on Hugging Face, which both companies have now addressed publicly.
Data Center Dynamics
AMD is putting up to $5 billion into Anthropic and deploying up to 2GW of MI450s via its Helios rack system, first gigawatt due H1 2027.
Data Center Dynamics
Contracted future spending jumped roughly $500 billion in a single quarter, and cloud revenue beat estimates without settling the cash burn question.
Electrek
Record revenue could not offset a profit miss and negative free cash flow, and the call went worse when Musk repeated familiar robotaxi and Optimus promises.
The Next Web
New research finds AI recommendations suppress verification behavior, producing worse answers held with more conviction.
