maximizing novelty and fun
United States
Joined November 2024
- Tweets7.1K
- Following405
- Followers189
- Likes41.3K
MacBooks are now so damn good I’m still using my M1 MacBook Pro (not the M1 Pro) and I still don’t have a single reason to upgrade
26
I see concrete examples of this every single day.
I think there is way too much concern expressed about verification of AI output in everyday work due to a risk of factual errors.
It used to be true, but data on more recent models suggest they beat human experts, including PhD specialists, in finding facts and checking claims.
8
Pulkit retweeted
Introducing Claude Fable 5.2 and Claude Mythos 5.2, designed to tackle the most daunting technical challenges.
They’re capable collaborators on ambitious projects in many fields, and show improvements in areas like coding, reasoning, and vision.
Pulkit retweeted
I don't know what changed, but for some reason the stars feel within reach for the first time in my life. I've been dreaming of exploring since I was born, but no matter how hard I pushed towards it, there was always the sense that it was permanently out of reach for me.
Something changed. I think its the models, and the amount of people who share in my hope. With capabilities increasing, and individuals acquiring such cosmic power to create things beyond our wildest dreams, the idea that *I* might be able to captain a starship no longer feels like a pipe dream. It feels like something I can prompt into existence. It feels like something we can *all* prompt into existence.
We're going to do it. We're going to become an interplanetary species. We're going to find or create new races to meet, and we're going to meet them with love. We're going to fucking do it and there is no amount of fear or doomsaying that can stop us.
Teaching AI to build scientific models (Anthropic Research, Nov 6, 2026)
A guest post from Anthropic AI for Science grantees Jakob Macke, Daniel Gedon and Stefan Wahl (ELLIS Institute Tübingen / University of Tübingen). Their core claim: AI can now build and fit scientific models fast, but a model that merely reproduces the data gives you an answer, not understanding. What creates understanding is knowing the whole range of models and parameters that fit, and which alternatives fail.
The anchor story is Hodgkin and Huxley's 1951 squid-nerve model. Cambridge's computer was down for six months, so Huxley solved the equations on a hand-cranked calculator, which took over three weeks. To fit the data they had to raise one gating variable to the fourth power without knowing why, and guessed that several charged particles had to move before a gate opened. Decades later (after their 1963 Nobel), the structure of a related potassium channel turned out to have four identical pieces around a central pore. The model, more than 70 years old, still describes neurons from squid to humans.
The authors' point: the fourth-power insight only carried weight because simpler versions demonstrably didn't work. As Richard Hamming put it in 1962, "The purpose of computing is insight, not numbers."
They split modeling into three repeated steps: build (choose components and equation shape), constrain (fit parameters to measurements), interpret (say what it means about the real system). AI is entering all three.
1
38
Constraining. Hand fitting gives one set of numbers and hides all the others that fit just as well. Their analogy: a $120 group dinner bill is consistent with many combinations of orders. Bayesian inference returns the full probability distribution of workable parameters, but computing the likelihood for every combination is intractable for complex models (e.g., thousands of randomly opening ion channels). Simulation-based inference (SBI) sidesteps this: sample random parameters, run the simulator (about 100,000 times for Hodgkin-Huxley, millions for bigger models), then train a neural network that maps an output back to the range of parameters that could have produced it. It only requires that the model can be run forward, even models that are just rules in code. Training takes weeks, but inference on new data then takes seconds; it's fast enough to localize neutron-star collisions in time to point telescopes, and it's used in cosmology, structural biology and geophysics.
The distribution shows which parameters are pinned down, where several explanations remain, and where another experiment would help most.
Building. Historically scientists proposed a handful of candidate models from experience in an essentially infinite space. In the past two years several LLM-agent systems have appeared that take data plus a natural-language description, write model code, fit it, compare to data, revise the equations, and loop. But run an agent twice and you may get two different, equally good models; returning just one hides the rest. So they argue for applying the same distribution idea one level up: the AI proposes many candidate models, each is weighted by how well the data supports it, with models that need lots of flexibility to fit getting less credit, weak ones drop out, and the agent proposes variations of the strong ones.
Features every survivor shares are likely supported by the data; disagreement shows where evidence is thin and what to test next.
Caveat: these systems have mostly been tested on rediscovering known models, with sparse evidence on truly open problems, though newer LLMs keep doing better.
1
27
Interpreting. This is the hardest step to hand to AI because "understanding" can't be scored like fit. Hodgkin and Huxley noted a third or fourth power was needed, and that higher powers might fit slightly better but the "improvement was not considered to be worth the additional complication." Their equations also explained the firing threshold, why pulses share a shape, and predicted that a slowly rising current never triggers firing. That is what philosophers call an "intelligible" theory: you can work out consequences without running it. Proposals for AI: judge understanding by how well it predicts the effects of disturbing a system, or judge whether its outputs help humans understand. Example: an AI for designing quantum experiments that writes a short program generating a whole family of setups, so physicists can read the design principle directly. A distribution of models also makes predictions robust: if all fitting models agree on what happens when every sodium gate is blocked, trust it; if they disagree when only some are blocked, that's your next experiment.
Scale. Classic models are small because humans had to handle them. AI lifts that limit. A fruit fly brain has more than 100,000 neurons (the rendering shows 124,635 cells from the male CNS connectome) and we now have its complete wiring diagram; models built on it already predict activity across the fly's visual system. At that scale there is no single correct model ("All models are wrong but some are useful," George Box). The question becomes which properties all data-compatible models share, and the best experiments are those that rule out the most remaining models. Climate science (ensembles of simulations) and LHC physics (maps of which theory versions are excluded) already work this way.
Their conclusion: we don't want to go back to three weeks of hand cranking, but much of a model's insight came from building it and knowing what failed. To offer understanding, an AI has to deliver not just the model that fits but an account of which parts the data pins down, which remain open, which alternatives failed, and what new puzzles emerged.
20
AI for scientific discovery: How researchers use Claude (Anthropic Economics, Nov 5)
Anthropic's Economics team studied how researchers use Claude. Lead authors are Johannes Hermle, Ben Jones and Lukas Althoff, with several coauthors. Using privacy-preserving classifiers that don't keep the open-ended research content, they labeled 207,049 consumer-plan research sessions from Aug 2 through Sep 30, 2026: 128,914 from Claude (Chat and Cowork) and 78,135 from Claude Code, plus a supplemental sample from Claude Science. Sessions were mapped onto SciNet's taxonomy of 7,262 tasks across 34 fields and 318 subfields.
Which fields use Claude most
• Relative to how many researchers work in each field, formal sciences (computer science, math, statistics) and the social sciences show up more than you'd expect. Lab-heavy physical sciences and medicine show up less.
• On Claude, the largest shares of sessions are computer science at 11.9%, engineering at 8.4% and medicine at 7.7%. Session shares track researcher headcounts fairly closely (correlation 0.79).
• On Claude Code the skew is sharper. Computer science is 4.3% of researchers but 32.7% of sessions (correlation 0.62).
• AI and machine learning alone are 5.0% of Claude sessions and 15.2% of Claude Code sessions, versus 1.2% of researchers. Computational biology, linguistics, neuroscience and robotics also over-index.
1
1
144
What researchers ask Claude to do
• Anthropic sorted work into eight research categories.
• Data analysis is 33.7% of Claude sessions and 48.1% of Claude Code sessions, versus 27.6% of how researchers say they spend their time.
• Writing and communication is 32.0% on Claude versus 16.0% of researcher time.
• Data gathering is underused: 11.5% / 11.9% of sessions versus 28.6% of time, especially physical collection. Digital collection is less underrepresented.
• Ideation and hypothesis generation is only 4.2% on Claude and 2.1% on Claude Code, versus 7.1% of researcher time.
• The single most common task is drafting manuscripts: 18.6% on Claude, 11.1% on Claude Code.
Survey and Claude Science
• A survey of 528 life and health scientists who are heavy AI users matched those patterns. They report the most use for analysis and writing, but still rank ideation first and data gathering second for producing important results.
• On Claude Science, which wires tools and databases into one workflow, biology and biomedical work are more than half of sessions, and ideation rises to about 8%, roughly matching how researchers spend their time. One example: proposing mechanistic models for how proteins interact with each other or with small molecules is 5.6× more common than on regular Claude.
• In biology and biomedicine, Claude Science sessions look longer (about 2× human hours), more doctoral-level (>80% versus about 50–70% elsewhere), and more autonomous (>70% high autonomy versus 23.7% on Claude and 64.1% on Claude Code).
1
1
50
Capability
• They scored each task by having an LLM compare research steps pairwise, then showed that Claude use tracks that score.
• Tasks in the bottom 20% of capability, such as packing a chromatography column or observing people in the field, barely get used. Tasks in the top 20%, such as visualizing data or applying statistics, get about 2× expected use on Claude and about 4× on Claude Code.
• Capability correlates with how well a task suits reinforcement learning (0.78) and with cognitive work (0.69). It correlates negatively with physical work (−0.57), interpersonal work (−0.31) and regulatory limits (−0.38).
• Subfields with higher average task capability see more Claude use per researcher (0.45 Claude / 0.57 Claude Code), and autonomy rises with capability too. Health underuses relative to its scores (regulation); formal sciences, social sciences, and arts and humanities overuse (they're friendlier to text and code).
Acceleration scenarios (illustrative; AI agent hour $30 vs researcher hour $95; AI runs 0.2 hours per researcher hour saved)
1. Current capabilities, wide adoption. AI takes 27–52% of researcher time (median 38%). Acceleration about 1.35–2.02× (arts/chem low; CS 1.96×, stats 2.02×). Physical share (~38–41% in lab fields vs ~4% median) explains ~⅓ of AI-share variation.
2. 75% of non-physical steps automated. About 8× CS, ~7× stats/LIS, ~6× economics; bench/medical ~2–3×; arts/history/anthropology ~2× or less.
3. 100% non-physical automated. Cognitive-light-physical fields ~15–16×; biology/chem/materials/ag/biomed stuck at ~3–4× without automated labs. Cost, not capability, becomes the ceiling (~16×).
What researchers expect. More than half expect AI to match top researchers at data analysis by 2027 (38% say it already does). By 2030: 94% analysis, 89% writing, 72% ideation and data gathering. About 40% expect full end-to-end AI research projects by 2030, about 64% by 2035, 16% never. Major autonomous breakthroughs mostly after 2035, if ever. Discussion flags slow diffusion, physical and ideation bottlenecks, quantity vs quality, organization/biosafety risks, and mixed job-satisfaction effects (e.g. US PIs report ~44% of research time on admin).
anthropic.com/research/how-r…
27
About half a year ago I built my first app, used it for a week or two and then completely remade it so that it works inside grok bot. The app is completely useless with how much AI has advanced, but it was a good problem solving exercise nonetheless.
It’s a simple RSS feed app which can track websites even if they don’t have a feed.
These days Grok Bot sends a detailed summary of the articles I want to read every morning. I’m also going to ask Grok Bot to post that morning briefing on my X account so you can enjoy it too.
The bot will post 1 summary a day, starting from latest -> old. So you might read something that happened a long time ago.
2
1
34