Michael Witbrock

@thsottiaux Treat control flow represented in python code and control flow represented in "data" as part of the same system when reasoning about code intent.

What is something that you feel is surprising that Codex still can't do well and we should have gotten right a while ago?

@pmddomingos Already happened in much of the academy, and at least AI may learn. Actually, it's not clear that being rewarded for it ever had much if anything to do with causing the production of knowledge.

The AI gap looms: when humans don't produce knowledge because they're not rewarded for it anymore, and AI doesn't either because it doesn't know how to yet.

For the very few who will care, in a relatively rare occurrence these days, Gemini (or whatever is behind "AI Mode" confabulates when it comes to David Hume's typographic practices in 1758. archive.org/details/bim_ei…

@IdoAizenbud @jakelevi_ml But this seems to be a poor mechanistic model of the actual biological neuron training context. Likely to help us find a good one though.

I think yes, but with an important caveat: larger point-neuron models may be good functional approximations, not necessarily good mechanistic models. If one dendritic neuron is replaced by many point units, the model may reproduce behavior while losing the mechanism. Then the “neuron” in the model is no longer aligned with the biological neuron.

What can a neuron compute? Real biological neurons are complex, but how capable are they? Using a new method, we found that a single cortical neuron can classify cats vs dogs, recognize spoken words, and solve 10-bit parity, all tasks thought to require entire networks. (1/15)

What isn't immediately clear to me is whether the training is biologically plausible. It's interesting but not particularly surprising that you can train a simulated neuron to solve complex tasks using backprop. What won't be clear until I read the paper is whether you can do so using plausible training contexts for real neurons.

One important point I didn’t emphasize enough: TwinProp is not just a method for neurons. The key idea is general: if you can train a sufficiently accurate differentiable twin of a complex object from input-output data, you can use that twin to propagate useful gradients, even when the original object itself is hard or impossible to differentiate. In our case, the object was a highly detailed spiking biological neuron. But the same principle could apply more broadly, for example, to optical systems, chemical reaction networks, fluid flows, various biological tissues, or any other physical systems where we can collect input-output data.

@lyft @RideBluebikes if you're going to make SMS to a US phone number the ONLY way to sign up, you really should consider sending the codes instead of lying and saying that you have.

@AnthropicAI Ah - that's why. To build trust, it would be helpful if the in-context message said this, instead of just saying it's not available.

The US government, citing national security authorities, has issued an export control directive to suspend all access to Fable 5 and Mythos 5 by any foreign national, whether inside or outside the United States, including foreign national Anthropic employees. The net effect of this order is that we must abruptly disable Fable 5 and Mythos 5 for all our customers to ensure compliance. Access to all other Claude models is not affected. We apologize for this disruption to our customers. We believe this is a misunderstanding and are working to restore access as soon as possible. Read our full statement: anthropic.com/news/fable-myt…

Is the Apple Macintosh charging brick the least-well-designed device ever? It blocks adjacent sockets completely. That is, until it falls out of the wall.

@_philschmid These are neither exhaustive nor exclusive possibilities. The usual answer will be both, dependent on context. The correct response here would have been no response to an unjustified dichotomy, but I'm only human. We should adapt both models and harnesses to avoid this in AI.

My personal research question for today: Should we optimize the model for a harness or should the harness be optimized for the model?

@pmddomingos Mops? Moles of ops.

The proper unit of computing is Avogadro's number of ops, by which standard it's still early days.

New Rules: processes that are compulsory ought to work. @WeWork Merchant Square London

@karpathy @shreyansj Are senior management "members of the administrative staff"? The MTS philosophy sounds good and in the case of Bell Labs was, but can also act as a way of homogenising and disempowering technical contributors. Look whether it applies to the parts of the company that imposed it.

Anthropic did not seriously make Karpathy "Member of Technical Staff"

@elonmusk @SpaceX @HiltonHotels Space, that is.

I need to go to the pace station in a PanAm space plane. @elonmusk and @SpaceX - please do the deal! @HiltonHotels needs get involcrd too. #late2001 cnn.com/travel/pan-am-…

I'm proud we are releasing LAION-fMRI, a densely sampled 7T fMRI dataset of natural images, with very broad stimulus sampling for testing countless hypotheses & deeply exploring brain representations. It is now available at laion-fmri.hebartlab.com What does LAION-fMRI offer? 🧵

Check out my latest article: Fast Loop: Aotearoa Agentic AI linkedin.com/pulse/fast-loo… via @LinkedIn

@OpenAIDevs I have many credits and I just bought more because of this message. Now take it away and let us get back to developing General Neurosymbolic AI.

@AsteroidEnergy What I said did not even begin to imply that. Perhaps @grok can explain why for you. Of course chemistry, physics and biology support useful crafts.

Computer science will stop being taught as a useful craft; it has become an explanation of how the world works, like physics, biology and chemistry. This change is long overdue.

@karpathy @karmicoder @jenzhuscott Yep, exactly recordism.org - a (slightly unsatisfactory) form of eternal life that you can actually have. And the technology is finally here to build it.

@karmicoder @jenzhuscott No the company physically gets you to come in, sits you down and asks you a precise set of maximally informative questions specifically designed for the best upload fidelity.

As I build my own 2nd brain 🧠 on Obsidian using @karpathy ‘s wiki idea, it suddenly dawned on me - one day when we r gone, our kids could inherit an interactive map to your mind, passion, obsessions, work, fascinations… It’s kind of beautiful way to think abt your 2nd 🧠.