given all the interest (thank you all), @nicochristie and i will build a tennis coach mvp next week, and thinking about golf next. pls leave a comment below if youd like early access we just processed and provided diagnostics and tips for the the first 7 players that have sent videos through. see here: was fun and learned a couple things. player A (personal friend) wrote me: "I like it. It matches what my coaches have Tried to tell me about my forehand, but maybe this distills it to simpler fixes and language that I’ll actually change" some more technical thoughts below - camera angle really needs to be in the middle of the court for "shoulder turns" and rotational metrics to be accurate - for juniors, some of the advice for shortening swings and such is probably ill-conceived, given that they have less power and are also still developing - player box and joint tracking, as expected, already work great across broadcast and phone footage, but stroke labels do not, so i had to get opus to manually double-check some of these. just set up a self-improvement loop to train a small dedicated classification model for this

It was shocking for me to see the current state of radiology scribe software. We watched extremely intelligent and trained individuals being forced to use 30 year old SEGA controllers to make changes. It honestly reminded me of macaque motor control experiments with joysticks back in neuroscience wet lab at Columbia. It was such a treat to work with real radiologists to build a realistic, diverse, and rigorous medical benchmark. Certainly an unfair advantage when radiology runs through @nicochristie’s family tree. We had originally assumed that hillclimbing this would be trivial, but there is a lot of fun and nonstandard optimizations that one can make when speed AND accuracy are both prized. In fact, in the month that we had worked on this, we got bitter lessoned twice (generalist multimodal LLMs are better at technical transcriptions than specialized STT models, and then Opus 5.5 not giving a shit about how bad the transcription is provided it is somewhat phonetically intact) Starting to learn about RIS and HL7 etc. Reach out to us if you’d like to chat, partner, or integrate this into your practices

The navigational system of the fruit fly is a crown jewel of systems neuroscience due to the work of some seminal neuroscientists (Larry Abbott, Gaby Maimon, Vivek Jayaraman, Barbara Webb), but it is an unfinished story. A fly that leaves a drop of food and wanders in the dark can always find its way back. To do that it has to keep a running sum of every step it has taken, a process called path integration. The neurons that report each step are known, but the neurons that add the steps up have never been found. There are two ways a brain can hold a sum like that. The usual answer is that some neurons holds it in activations and sustains these activations by exciting each other in a loop tuned so precisely that the signal neither fades nor blows up. Most models of navigation assumes this, and it is how RNNs and LLMs hold state too. The other answer is that nothing keeps firing at all. Each step is encoded into synaptic strengths (aka weights), and the sum of all synaptic strengths is the sum of the journey. A few papers have suggested the fly works this way but nobody has pointed to any candidate neurons, until now. Four neuron types, hΔH, hΔA, hΔI and hΔG, have no known functions, but we found that they have every ingredient option 2 needs. They receive input from neurons that report each step taken, they receive velocity-sensitive dopamine input that could gate memory writing, and they receive a reward-sensitive octopamine neuron that could reset the synaptic weights at food arrival. Simulations confirm this is a viable candidate for path integration. This is the key finding, but we have posted 3 other findings in links below. We tried to be exhaustive with published papers but may have very well missed some key published results, so inviting the cogniscenti to engage. Background, methods, experiments, results, as well as relevant citations: Panoramic circuit view: GH repo:

I and @nicochristie ran the fly connectome and found the group of neurons (hΔH, hΔA, hΔI and hΔG) that could allow the fly to navigate using fast synaptic weight updates, not neural activations. This is fast-weight continual learning in a fly, something current LLMs don't do!

whenever something is crazy hype and is followed up by an armada of PR, you gotta bring out the bullshit detectors so i cloned the repo and did a little digging and found a lot of sound and fury. please enlighten me if any of these facts are untrue, would love to be educated RLM claim either overblown or false this doesn't even appear to be RLM ( RLM is "cool" as a concept because there's a premise of going infinitely deep to do arbitrary decomposition of a problem. in practice though it's not really hard, you just make subagents callable as functions and wrap that in a code execution tool. the reason RLMs have not been productionized and is just seen as a research ditty is that 1) going arbitrarily deep down the stack is a shitty thing to do in production without guardrails, 2) most problems do not need more than 2 layers of decomposition, and 3) agents are bad at banana phoning each other and will fail to preserve the subtleties of human requests, leading to a lot of inefficiencies and bad behavior. to solve these problems, there needs to be a fair bit of innovation, either on model training end or on harness end, to mitigate these issues. so what were the innovations in prime agent? they set RLM_MAX_DEPTH = 1? ... ok what? so theres no innovations, it's just calling subagents like every other harness in existence. if RLM MAX DEPTH is 1, it's not even RLM? it's just... a harness? so now i try to look up to what depth the ARC AGI 3 hillclimbing was done at. not surprisingly this doesnt seem to be disclosed on the blogs / posts. if it's just 1 layer deep, that's just shady marketing. a real innovation or contribution to harness engineering would be to detail the things that were improved to get a max depth of >3 to work without blowing up your computer, wasting a bajillion tokens, taking forever, or having agents go off the rails. i didnt find any 97% is just overfitting after hillclimbing on a public eval ok folks. it's nice that opus5 does better than sol than terra, but at the end of the day, we're looking at a public eval in which problems and solutions are open access, so any agent harness can arbitrarily overfit to the task set however much it wants to. if you do a codex or claude loop with some semi shady prompting to just get 100%, im pretty sure it will too. now obviously this is almost certainly not what the prime intellect team instructed their agents to do (otherwise you'd see that terra get 100% after a couple iterations), let's just all remember that there is no train / test split, and that ARC explicitly says - Public-set scores are vulnerable to task-specific overfitting. - They are “emphatically not” valid evidence of progress toward AGI. - Real generalization should be tested on the 55 semi-private or 55 fully private games. It's difficult to assess the degree of overfitting in Prime Agent's self-improving harness (which, btw, is not terribly different from nous agent or really any self-improving system that just looks at old JSONL agent trajectories and suggests skills / memories / system prompt appendments / tools / extensions). the more overfitting, the less impressive the 97% performance is. a counter-example that i found to be legitimately impressive was -- minimal / almost no overfitting, stupidly simple and general solution, and 95% on ARC AGI 3. i think the fairest thing to say about this system, given it's independent daemon system for managing subagent lifecycles, is that Prime Agent contributes a fairly robust persistent asynchronous agent-process tree. It does not demonstrate a solution to scalable deep recursive agency