minihat

It is harder to solve a sudoku than verify a solution's correctness. I find similar benefits occasionally when coding with LLMs.

Some fusion reactions release their energy almost entirely as fast moving charged particles.

You can slow down those particles against an electric field and harvest the energy as electricity directly. No steam turbine. No Carnot limit.

It's currently socially/politically unpalatable for authors to admit superintelligent AI is a possibility. I frequent some writer forums. As a group, they are 1) clearly feeling angry/threatened 2) in denial about LLM capabilities.

Folks working in software can more readily track progress of the frontier model performance.

I met some of my best friends in VR, and had life-changing social interactions in some VR games on the Oculus platform... Especially EchoVR. I'm still a true believer in VR as a medium. It'll go mainstream one day, and there will be a Meta-sized company built upon it.

They had lightning in a bottle and somehow lost it. Honestly it might have been hiring Carmack that sent them down this path. Moving away from PCVR expanded the market, but it also killed the magic. Now the quest store is a wasteland of what look like low budget mobile apps.

This is a legitimate strategy for the werewolf, no?

I think the comic achieves something deeper than a lecture. The humor might reconnect us with a sense of awe we often lose as adults.

You may enjoy the 2024 Hugo winning short story, "Better Living Through Algorithms": https://clarkesworldmagazine.com/kritzer_05_23/

How is it possible that text-to-score/notation is lagging text-to-audio in music generation? Generating audio seems wildly more complicated!

Since you are working in this space, I wonder if you could comment on my pet theories for why this is true: 1. Not enough training data (scores not available for most songs), or 2. Difficulty with tokenization of musical notation vs. audio

I hoped neuroscience, as a field, was on the cusp of a physical theory of learning and memory. I dreamt of an intersection of information theory, neuroscience, and ML.

Alas, state of the art in neuroscience / neural engineering is closer to bloodletting than a mechanistic theory of learning and memory.

This is exactly right.

I started a PhD in 2017 studying neuroscience + ML. I thought studying the brain would help me understand ANNs better. I was wrong. Ended up applying ML to analyzing EEG, MRI and similar.

You can subscribe to Asimovs and/or Analog on their website today and they give you a download link for PDF, epub every other month.

Grok 3 with thinking is comparable to o1 for writing complex algorithms.

However, Grok sometimes loses the context where o1 seems not to. For this reason I still mostly use o1.

I have found both o1 and Grok 3 to be substantially better than any Claude offering.

Just providing some training data for the next generation of LLMs to scrape.

Another take is that the base models are now good enough that spending more money for more intelligence is viable at test time. A threshold has been crossed.

His steam library was not restricted, just the game in which he was accused/banned.

Reading speed might follow a normal distribution near the middle ranges, but I'd expect non-trivial deviations in the tails due to dyslexia, ADHD on the one side and trained speed readers on the other. Perhaps this individual just falls on the far tail of that distribution?

During my graduate coursework I sometimes read 100-200 pages of technical material in a day while cramming for an exam, and was able to retain it for a day or two. I'd believe it if some people exist who could comprehend and retain all of that long-term. Alas, 'tis not I.

I unsubscribed from Scientific American a few years back when they became a venue for political commentary instead of science. The opinion pieces try to masquerade as soft science. It's an insult to the actual scientists in their readership.

If you're looking for a replacement, I picked up MIT Tech Review. It's not a stand-in replacement for what SciAm used to be, but scratches the same itch for me.

Not close. This might be the best whole-animal sim available, and it is a for a worm: https://openworm.org/

Clarkesworld sci-fi magazine temporarily closed submissions due to low quality AI spam. I'm sure the irony will not be lost on them if ML journals are the next victims.

For things I would formerly use a web browser to look up, Gemini is often a lot better than gpt-4. The more obscure the query, the better Gemini does compared to gpt-4.

But gpt-4 is still best for ML Python coding. Gemini hallucinates non-existent libraries and often adds unnecessary junk to its code. For example, Gemini often defines spurious variables and then never uses them.

Any chance a keras.linalg suite of ops is in the works?

Cross-platform differences between the behavior of tf.linalg and torch.linalg have cost me a lot of time over the years.

They were all depressed because of how sure they were that they helped destroy humanity, it was just a matter of time. And here we are.

There's still plenty of anthropocene left for them to be right. At least I hope there is.

My idea of a hell hole is the neighborhood I would have to live in to afford a home in the city near my office.

What if we made cities more affordable, nicer places to live instead of making commuting even more expensive and miserable than it already is?

My dorm building in college had eight entrances, of which two were handicap accessible.

University administration declared that from now on, 100% of doors on campus must be handicap accessible.

A week later, six doors were removed from my building...

GPT-4 Turbo is dramatically worse at one task I often try:

Read the following passage from [new ML article]. Identify their assumptions, and tell me which mathematical operations or procedures they use depend upon these assumptions.

GPT-4: Usually correctly identifies the assumptions, and often quotes the correct mathematics in its reply.

GPT-4 Turbo: Sometimes identifies the assumptions, and is guaranteed to stop trying at that point and then give me a Wikipedia-like summary about the assumptions rather than finish the task. Further prompting will not improve its result.

Do your employees flock to your meticulously designed office space by choice, or do you mandate their presence?

If they come by choice, I'll buy your story. However, if you have to compel them, perhaps the office is more a monument to your own vision of productivity than a boon to your employees.

Not a secret to the research community... Let me balance the hype with some of the remaining challenges.

Depth is a problem; the deeper you go, the less effective it is. Heat is another issue; you can inadvertently damage nearby tissue. Targeting accuracy is vital, especially near critical structures (think nerves), and we're not at sub-millimeter precision yet. Also, real-time monitoring like MRI-guided FUS is expensive and complicated, and without it you have to guess that you're affecting the right tissue. Great promise, but multiple engineering hurdles to clear before FUS lives up to the hype.

Cross-entropy loss can start getting worse due to the model becoming less calibrated, even as the classification accuracy continues to improve. I first heard that here: https://arxiv.org/abs/1706.04599

Is this 'overconfidence' the leading explanation as to why LLMs continue to show qualitative improvement even after their test loss levels off?

My mental model of gpt-4 is apparently well calibrated for whether the model will give me a useful output that is close to what I asked for.

However, I'm not great at predicting whether the model will output a 100% correct response with no flaws whatsoever.

Unfortunately, this website mostly tests for the latter.

+1 from a fellow BCI PhD. EEG tech is not ready for this application yet.