BAI Research

@DefenseAnalyses They are more likely to be aligned with the US in the future.

Civilizational allies require civilizational nations. America should welcome the rise of nationalist parties in Europe even when they are notionally adverse to us. In fact, it is this credible assertion of independence that in part creates conditions for European revitalization.

@Nature Human cooperation is largely a function of their inferred intra vs inter population genetic distance, as the species is adapted to optimise its long-term indirect fitness through group coordination (e.g. language).

Nature research paper: Punctuated decline of human cooperation go.nature.com/4toTIvn

@CoreAutoAI A implies B

Deep learning works

@andrewchen This is a beautiful experimental design; one can retrospectively measure to what extent a past event (or invention) could have been predicted based on historic data.

so it’ll turn out that psychohistory from Foundation was a diffusion model trained on many date cutoffs then able to extrapolate the future? How it’d work: 1. **Train vintage models at cutoffs** e.g. pre-1931, pre-1950, pre-1970, pre-2000, pre-2020. 2. **Ask them to predict/discover the future** Can the 1930 model infer nuclear weapons? computers? DNA? information theory? modern geopolitics? Can the 2000 model infer transformers? crypto? COVID-like pandemics? AI scaling? 3. **Score “future-predictive latent structure”** If a model repeatedly extrapolates correctly beyond its cutoff, that suggests the historical data already contained compressed causal structure pointing toward later discoveries. 4. **Apply the same setup to today** Train on pre-2026 data, ask for 2027–2035 predictions/discoveries, then sample many candidate futures. The diffusion-ish part is: instead of denoising pixels from noise, you’re “denoising the future” from the latent causal constraints in the past. History so far is the conditioning signal; possible futures are samples from the posterior.

@benedictk__ There is a serious discrepancy in the performance one can obtain based on how people use these models (Codex included).

I thought it was common knowledge to have gpt image make UIs and then have 5.4 or 5.5 build it? Same for slides actually. For real do people still do the click thing and not just one shot their slide deck using image models?

@PNASNews Does the thesis hold if the cartoon characters are the same colour, and thus signal genetic proximity?

One-year-olds watching cartoon characters help each other don’t necessarily expect a helper to also help new characters, suggesting that infants assume prosocial behavior is based on relationships rather than innate dispositions. In PNAS: ow.ly/hjep50YNskl

@OpenAIDevs @TheRealAdamG There is opportunity to develop more robust search tools for both chat data and Codex session logs.

Introducing Euphony, an open-source tool for visualizing chat data and Codex session logs. Paste in a public URL or upload a local file, and Euphony turns the raw data into an easy-to-browse view. It supports translation, filtering, editing, and more.

@ScienceMagazine The presupposition that group cooperation necessarily conflicts with individual interest is derived from WEIRD ecologies where there is no genetic interest for individuals to cooperate with their group relative to external competing populations.

Punishing freeloaders in public good games—experimental models used to analyze the social dilemma between individual self-interest and group cooperation—can boost cooperation, but whether punishment helps or harms the groups’ outcomes depends heavily on context, according to a new Science study involving over 7000 human participants. The findings reveal when, rather than whether, punishment works. scim.ag/4e9YVSY

@ScienceMagazine Such interpretations cannot be granted without measuring the genetic distance of the factions and comparing it to the genetic distance of other populations typically associated with intergroup conflict.

Chimpanzees regularly fight viciously over food, mates, and rank, but only rarely do these brawls spill over into a broader civil war. Now, a study tracing 30 years of chimp behavior in Kibale National Park in Uganda reveals how and why such internecine violence erupts. The study shows how in chimps—and perhaps humans—tensions in once-peaceful groups can grow into deadly violence, even without resource shortages or cultural divisions to fuel them. Learn more: scim.ag/4dJJrFd @NewsfromScience

@Britannica *Encyclopedic discretion when republishing interpretations about what could be a natural population emergence event. Genetic distance measurements have not been reported between clusters; how they objectively differ from populations commonly associated with inter-group conflict.

Chimp. Civil. War. This is not a ‘Planet of the Apes’ sequel, but a real-life phenomenon currently happening in Kibale National Park in Uganda. What do battling chimpanzees have to teach us about keeping the peace? 🧵⬇️

@ScienceMagazine Did they measure the genetic distance between the two clusters and compare it to the distances between populations typically associated with intergroup conflict? Seems to be a case of population emergence.

In a rare, decades-long study, researchers in Science have documented what appears to be the first observed “civil war” in wild chimpanzees. The findings demonstrate that shifting social ties alone can fracture a once unified group and ignite sustained, deadly conflict among former allies. scim.ag/4dDhoHq

@unconvAI Cool

We’re introducing the Unconventional Grant. A new research grant program supporting bold, unconventional ideas in AI. We’re allocating $500,000 in total funding, awarding up to five $100,000 grants to researchers exploring new paradigms in efficient, scalable, and biologically inspired AI systems. We’re especially interested in ideas that challenge how AI systems are built today, from unconventional circuits and architectures to new approaches in neural networks and theory. Not incremental work. Not safe bets. Ideas that push the field forward. unconv.ai/grant/

@jeffreyweichsel It is fortunate you are analysing a species that is uniquely adapted to optimise its long-term genetic survival through population coordination (language), meaning its behaviour can be predicted by calculating the lasting benefit conferred on its distributed genetic information.

WANTED: PSYCHOHISTORIANS Seeking exceptional candidates with advanced training in stochastic processes, sociology of masses, and predictive modeling of large-scale human behavior. Candidates must demonstrate: • Mastery of n-variable differential equations governing societal dynamics • Ability to extrapolate historical trends across millennia with statistical rigor • Willingness to operate under conditions of utmost discretion Successful applicants will join a small, elite team developing a revolutionary framework for forecasting the future of the Galactic Empire. Work involves collaboration with Dr. Hari Seldon and select colleagues. Compensation: intellectual fulfillment + secure accommodations on Trantor (or relocation to designated peripheral projects). This is not a position for those seeking public acclaim or imperial favor. Only those who understand that the fate of trillions may rest on precise probabilistic analysis need apply. Reply via encrypted channel: Project Prime Radiant – Reference “Seldon Initiative.” Vague inquiries will be ignored. Talent will be noticed.

@MillionInt Whether gradient descent on a leaderboard will approach a global maximum.

Say what you want, I think this was the top SF billboard. I think about it a lot

@benedictk__ @lateinteraction

@lateinteraction 🤣

everyone accepts Claude being called Claude, but there would be outrage if OpenAI was like “hey ChatGPT 5.5 is now called Walter”

@TimSweeneyEpic They need to restore the ability to set the taskbar interface to an always combine and show labels state (e.g. 7+ Taskbar Tweaker on previous versions).

Great moves by Microsoft!

@industriaalist The fundamental limitation with a parameter count constraint is that it overlooks activation sparsity; for example a model with 100x parameters could only have 1% active for any given input sequence and thus requiring only approx 1x compute.

few thoughts on openai's parameter golf: - first, you'd be surprised how many researchers at big labs (not just openai) are interested in our slowrun - i'd expect openai to be already automating parameter golf *entirely* with agents. and i'd also expect agents to be better than humans at this already. - for slowrun, we've deliberately kept it less gamified. the search space over learning algorithms for data efficiency is much larger than for compute/parameter efficiency. so slowrun is less of a competition and more of an open research effort toward interesting, new learning algorithms

@willdepue @kellerjordan0 The compute cost of model parameters is a function of the learning algorithm; to prevent bias of research towards conventional learning algorithms consider constraining input data or compute.

we're launching a challenge, inspired by @kellerjordan0's NanoGPT Speedrun, to train the best model under extreme parameter limitations anytime i get asked how to break into ML, i point to challenges like this: just start training models! also, we're covering $1M of compute