a new anthropic study says AI can already do about 38% of the work behind a scientific study: going through data, analysing it, writing it up
if scientists used it for everything it's good at, research in statistics and computer science would go about twice as fast
chemistry and biology gain less, because someone still has to run the experiments in a lab teapawt.net/emollick/status/211885…
Some new data from Anthropic looking at the impact of AI on the sciences by field: "With current capabilities, if adopted widely, AI could take on 27–52% of the researcher time needed to produce a new study, with a median of 38% across fields..."
Here: anthropic.com/research/how-r…
this week deepmind's AI agents found two new weaknesses in RSA, the math behind digital signatures and secure connections. one of them makes cracking many keys at once about 30x cheaper per key
it only works on old 1024-bit keys. how long until it's the 2048-bit keys' turn? teapawt.net/GoogleDeepMind/status/…
We’re sharing new research demonstrating how AI agents can accelerate the science of stress-testing encryption.
Guided by humans, our models uncovered two major weaknesses in legacy 1024-bit RSA cryptosystems.
Our VP Science @Pushmeet explains more ↓ teapawt.net/pushmeet/status/211875…
people don't realize light is slower inside glass. in a fiber cable it travels about a third slower than in space
and almost all the internet traffic between continents still runs through cables on the ocean floor
so the fastest cable between new york and london might end up being no cable at all teapawt.net/elonmusk/status/211878…
same technology, opposite results
in china, almost 27,000 students started using AI for homework. homework scores went up 18%. exam scores went down 20%
in el salvador, 171 public schools gave kids AI tutors. they now score about as well as kids in germany and sweden
AI that does the work for you makes you worse. AI that teaches you is the best tutor most kids will ever have teapawt.net/TheEconomist/status/21…
Widespread adoption of AI has fuelled concerns about the impact the technology might be having on learning. Our charts help explain its educational effects economist.com/graphic-detail…
everyone keeps saying we're not paying the real price of AI yet, and one day it'll hit us
i don't buy it. a year ago anthropic spent $1.72 on compute for every $1 it made. now it's 57 cents. the trend will keep on going
though if i'm wrong, get ready to learn rust teapawt.net/TheEconomist/status/21…
Earlier investors who had hoped for a public valuation as high as $2trn for Anthropic may be disappointed. We explain why economist.com/business/2026/…
i don't think people realize we're at the point where AI does math that's correct and the best mathematicians in the world can't follow it teapawt.net/TheEconomist/status/21…
In recent months AI labs have flooded the mathematics field with papers that are almost impenetrable to human minds, and kept the details of new results under wraps economist.com/science-and-te…
the big model was re-reading the whole chat every 30 seconds just to write a one-line status. now a small model does it from notes. good catch teapawt.net/ClaudeDevs/status/2118…
the rule that makes this work: whoever finds a bug can't be the one who confirms it. same reason you don't grade your own exam teapawt.net/CloudflareDev/status/2…
With the recent wave of vulnerability disclosures, it’s a good time to take a closer look at your own code.
Turn your coding agent into a team of security auditors with the open-source skill that seeded Cloudflare’s vulnerability discovery harness.
The 6 phase approach:
1. Map the codebase
2. Hunt for vulnerabilities
3. Challenge each candidate
4. Record the findings
5. Verify the findings
6. Report what’s confirmed, rejected, or still needs validation
Start your hunt here:
github.com/cloudflare/securi…
i remember when his name started showing up everywhere. ex-openai, wrote the "situational awareness" essay in june 2024, then his fund did +47% in the first half of 2025. reportedly $20B+ under management
i genuinely told myself: i need to follow whatever this guy does next
somewhere along the way i learned about the S&P 500 and how much safer it is long term. boring monthly DCA
one big win doesn't promise the next one
S&P 500 is the right choice 🔥 teapawt.net/MrMikeInvesting/status…
JUST IN: Leopold Aschenbrenners Fund (Situational Awareness) has reportedly officially blown up…
These are the current (now liquidated) holdings:
- $MU bought at $1,153, now $703, (-39%)
- $CORZ bought at $25.6, now $11.1, (-56%)
- $SNDK bought at $2,268, now $874, (-61%)
- $NBIS bought at $273, now $121, (-55%)
- $CRWV bought at $99, now $42, (-57%)
- $BE bought at $301, now $165, (-45%)
Use him as a lesson why you NEVER over-leverage into high risk positions.
That concludes an end of an era…
shopify rebuilt its mobile apps so an AI can use them the same way you do
every screen describes itself as plain data: what's on it, and every button you can press. an agent reads that and presses the buttons, no screen needed
they built it so their coding agents could test the app in milliseconds. and their AI assistant uses the exact same interface to do things for merchants inside the app
shopify calls it "agentic-first". the app has two kinds of users now: people and agents teapawt.net/ShopifyEng/status/2118…
We rebuilt our mobile apps in Swift and Kotlin with agents. To keep pace and hold our quality bar, we built Reflex: an agentic-first architecture that lets agents build and verify app behavior across iOS and Android without needing devices or simulators.
Here's how it works, piece by piece:
shopify.engineering/reflex
mozilla calls firefox "one of the world's most hardened and scrutinized codebases". open source, used by millions, bug-hunted for 20+ years, already scanned by several generations of AI models
GPT-6 Cyber still found 81 high-severity bugs nobody knew about, and showed 30 known ones were worse than reported
turns out the slop was human-written all along teapawt.net/OpenAI/status/21183901…
Introducing GPT‑6 Cyber: a major advance in security research that pairs stronger capabilities with better model alignment.
The model powered the Linux kernel research we announced yesterday, and security teams at companies like Mozilla, Palo Alto Networks, Calif, and Cisco—alongside our own at OpenAI—are already putting it to work.
GPT-6 Cyber is available to select Daybreak Red customers and we are expanding access over time.
openai.com/index/introducing…
40 years ago scientists found giant barrels inside our cells and still don't know what they do. a team just used claude to find them in bacteria and take a 3D picture of one in 34 hours teapawt.net/AnthropicAI/status/211…
We’re sharing how our biology research team worked with Claude to map the first 3D structure of a vault particle from bacterial life.
Vault particles are a longstanding mystery. They are barrel-shaped objects, among the largest protein structures found in our cells. They were first discovered by Leonard Rome and Nancy Kedersha in 1986, but their function remains unknown.
With Claude, we’ve found results that might bring us closer to solving that mystery.
Until now, vaults were only known in the cells of complex things like animals and plants. We have shown that similar proteins from bacteria (and the viruses that infect them) can also assemble into vaults. Finding the structures in these simpler forms of life may make it easier to investigate their function in our own cells.
For our life sciences team, the project was also a demonstration of the potential (and the limitations) of using AI to accelerate science.
In our full piece, we discuss how Claude was able to dramatically speed up some stages—manually determining the vault particle’s structure would have taken weeks to months of painstaking work, but here it took days. We also discuss how Claude made some simple mistakes along the way, and how we collaborated with our machine learning researchers to address them.
Read more: anthropic.com/research/inves…
i don't think people realize what happened this week
- a 2.8 trillion parameter model was released for free
- an AI edited another model's brain and made it 6x better at math without training it
- an AI found and patched 948 bugs in the linux kernel in one run
and it's only thursday
> 948 privilege escalation bugs in the linux kernel
> found, validated and patched
> in a single run
35 years of the best engineers on earth reviewing this code, and a machine found 948 holes in one pass teapawt.net/OpenAI/status/21181194…
Our security researchers built an autonomous cyber vulnerability discovery, exploitation, and patching pipeline.
In a single run with an upcoming cyber model, it validated and generated patches for 948 local privilege escalation vulnerabilities in the Linux kernel at a total API cost of $1.2 million.
the sixth most visited website on earth just rebuilt its web app, and the first version was built by 5 engineers with AI agents in under 6 months
on a slow connection the old site showed the X logo for 22+ seconds before the first post. the new one shows it in 2.7
the old setup: 3 different servers (one for people, one for google, one for link previews), each building the same page its own way. any SEO fix had to land 3 times
the new one: 1 server that checks who's asking
- people get HTML streamed in as the data arrives
- google gets the full page in one piece
- link preview bots (imessage, slack, discord) get only the title, description and image
how 5 people did it with agents:
- the old servers became the spec. an agent reads how the old code handled a page, ports it, and writes a test that pins it down. a script requests the same URL from old and new as googlebot and compares the results
- one AGENTS.md file with the house rules: named exports, no barrel files, validate outside data, nothing gets committed until lint, typecheck and tests pass
- 25+ skills in the repo for recurring work: commits, stacked PRs, code review, getting a PR through CI, translations
- an "anti-slop" lint plugin that rejects the patterns agents love: chained type assertions, parameters typed as unknown, conditional spreads of empty objects
the results:
- 0 bytes of javascript downloaded before the first post shows (the old site needed 3+ MB)
- files loaded at startup: 754 → 182 signed in, 582 → 10 logged out
- 36,000 lines deleted with one old server
- one server fleet cut from 1,400 to 700 instances per data center
the classic advice is "never rewrite from scratch", because the rewrite stalls while the old product keeps moving. with agents doing the implementation it took months, not years
the engineers' job became deciding what the system should do and checking in a real browser that it does teapawt.net/yagiznizipli/status/21…
The web rewrite is working.
On a slow connection, the old X web took more than 22 seconds to show the first post on a profile. The new X web does it in 2.7 seconds. Here's how we did it, and how we're rolling it out one country at a time.
blog.x.com/en/engineering/re…
> touched 0.016% of the weights
> math up to 6x better
> no training
the models are starting to do brain surgery on other models and people are still asking if AI is a bubble teapawt.net/rikiyatakehi/status/21…
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.
It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.
We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵
Blog: rikiyat.github.io/blog/weigh…
Readers added context they thought people might want to know
LLM was given weights explicitly as text in prompt and used code/forward passes (per authors' logs) to ablate Qwen looping direction, a known thinking-mode issue per model card. Primarily fixed repetition in that mode, not general 'natively' improving weights/math.
rikiyat.github.io/blog/weight-sp…
github.com/RikiyaT/lm_und…
huggingface.co/Qwen/Qwen3.5-0…
huggingface.co/RikiyaT/Qwen3.…
It's Halloween, so I made a zombie shooter you can play in your browser 🎃
- the dead come by wave
- one double-barrel shotgun
- headshots count double
Built in a 2.5 hours with three.js + Claude Code
Play: tokelde.com/games/sky-horde
Heads up! Things fall from the sky.
Don't get squashed.
Top survivors make the world ranking.
How long can you last?
Built with Opus 5.5 + three.js
Works on desktop and mobile.
👉 tokelde.com/games/heads-up/
Added wildlife to my kayak ride: race a pod of porpoises (Shift to sprint), gulls dive for fish, and a sea eagle snatches one off the water.
three.js + Opus 5.5
Runs in your browser
Paddle out: tokelde.com/games/kayak-fjor…
I built a kayak ride through a Norwegian fjord that runs in your browser.
three.js + Opus 5.5, 3 hours, 195KB gzipped. No textures, no models: every cliff, tree, wave and the northern lights are generated in code.
Golden hour to aurora, every reflection computed live.
Grab a paddle: claude.ai/artifact/NQCWfRzvp…
