tbenst

Put another way, EEG moves top-10 retrieval out of 512 candidate words from 20.3% -> 27.7%. Hardly "open-vocabulary word decoding"

The paper is also missing a critical control of feeding Llama the preceding ground-truth text and asking how well it ranks the current word among the same candidate set.

That is sobering for sure, I wonder what the theoretical bounds are on what is possible if known. Would be such a dream to use a Frontier LLM one day with homomorphic encryption, but this sounds wildly implausible based on where things are today.

A key challenge with Alzheimer’s is there is no good mouse model for the disease. While some approximate the phenotype, it’s not clear that the disease model as commonly studied in mice matches well with mechanisms of the human disease. There’s some thinking in the field that this could be a key reason why so many treatments have appeared very promising in mice and haven’t panned out in humans.

Beautifully written and worth the read. And the screensaver nerd snipe is epic.

As a neuroscientist, my biggest disagreement with the piece is the author’s argument for compositionality over emergence. The former makes me think of Prolog and lisp, while the later is a much better description for a brain. I think ermergence is a much more promising direction for AGI than compositionality.

You’re right, I meant class 4

This is a well known phenomenon. It accounts for example in the flash perceived when someone inadvertently looks at an infrared class 5 laser and is blinded

This is fun, and the modeling is cool for sure, but it's well known that ultrasound can be used with surgical precision in the human brain.

Focused ultrasound is already used for non-invasive neuromodulation. Raag Airan's lab at Stanford does this for example using ultrasound uncaging.

https://www.frontiersin.org/journals/neuroscience/articles/1...

https://www.sciencedirect.com/science/article/pii/S089662731...

Also see the work by Urvi Vyas, eg

https://pubmed.ncbi.nlm.nih.gov/27587047/

I don't mean to discount the cool imaging-related reconstruction of a point spread function, but rather to say that ultrasound attenuation through the skull an soft tissue has already been well characterized and it's not a surprise that it is viable to pass through.

The Nvidia GH200/GB200 “superchips” are all ARM processors. Seems likely that some of the next generation of foundation models will be trained on ARM

I know the exercise was to p-hack, but instead I decided to one-shot my attempt at the most reasonable model from first principals:

- given that we are looking at a national scale, use only national politicians

- use the components from Macroeconomics 101: exclude inflation as that’s on the Fed, exclude stocks as too conflated with FX and international investing alternatives

- don’t needlessly withhold data

Tried one hypothesis, so p-value of 0.04 is accurate. Still OK to explore if you Bonferroni correct the p-Val afterwards

Any references for 99.9% density with SLM copper? My understanding is that pure copper SLM printing is less frequently done as doesn’t work well with the infrared lasers on most machines, requires high heat & speed, and has more porosity than other alloys. It’s also hard to print so that it’s strong, conductive and heat stable.

I think there’s still quite active research in the area, though, and no doubt there’s a lot going on that I don’t know! https://www.sciencedirect.com/science/article/pii/S026412752...

https://www.ncbi.nlm.nih.gov/pmc/articles/PMC10861549/

The high thermal conductivity of copper makes it difficult to maintain needed temperatures during SLM. Also, copper is prone to oxidation at high temperatures, further complicating (thermal based) laser melting 3D printing techniques. It’s more typical to print copper alloys than pure copper.

Fabric8Labs can print 100% density, whereas Desktop Metal is highly porous. Also Fabric8Labs can directly print pure copper, which has historically been very difficult. The process is also more energy efficient and better suited for small complex parts. Desktop Metal serves a different market in terms of material and size.

disclaimer: I'm a GP at Asimov Ventures and invested in Fabric8labs' pre-seed round.

Our preprint does exactly that :). https://arxiv.org/abs/2403.05583

x1798DE captured my intent well. For example, tonal languages like Mandarin or Cantonese may be more difficult to decode if vocal cords aren’t vibrating, and languages with more phonemes that have both a voiced and unvoiced version might be more difficult. I still think decoding will be possible for general language, but that’s a hypothesis whereas I know it’s true for English.

This is a super cool device. Note that the decoding is highly limited: they decode into one of five different sentences. This is easier than five words for example as there is more information to distinguish.

Unfortunately the media is blowing this way of out proportion as the larynx alone does not contain sufficient information to decode silent speech.

If you also sense the lips, tongue articulators, and jaw, then general English decoding becomes possible with high accuracy (eg see our recent work here: https://x.com/tbenst/status/1767952614157848859). It’s not in the preprint but I’ve done experiments with only the larynx recorded and performance is pretty abysmal on even a 10 word vocabulary—-hence why they did a five sentence task.

Very interesting but hard to interpret until the performance numbers / benchmarks are available. I can already fine-tune a 70B language model at home using CPU + RAM, but it would be so slow as to be almost totally impractical (~20x slower than GPU). It would be great to see a comparison to eg 8 x A100 (available for $32/hr on AWS on-demand) and also CPU + RAM. Presumably it’s somewhere in between, but hard to predict where!

There’s an amazing effort by a nonprofit to sponsor the first Kayak descent of this river by children from indigenous peoples of the region. This is a seriously intense adventure with anticipated class VI rapids.

https://www.riostorivers.org/

Does anyone know the state of running Windows / Linux x86-64 virtualization on Apple Silicon? This article is super interesting but dances around the most important application for VMs on Mac.

It’s more about the difference in magnitude of the lenses. This also gives you depth information when combined.

Not sure he can claim the phrase. On cursory glance, can find the phrase in multiple books prior to the publication of the cartoon, using the same phrase. Seems it was already in the vernacular.

1986: "How about the afternoon?" "How about never?”

https://www.google.com/books/edition/Wild_Nights/IaAA0LV0zWk...

1981:

there had to be someone else in the whole school he could talk to besides her. "How about never?" She turned back to her notes and didn't even notice when Harry left the room.

https://www.google.com/books/edition/The_New_Voice/4pQSCfcy4...

Thanks for the response! So does vscode.dev, but it doesn’t support “Remote - SSH.” Could you confirm if this specific extension is functional?

To expand, for Remote - SSH to fully function it must read a ~/.ssh/config file as well.

The web version you linked does not support the remote SSH extension, which is critical for many folks workflows. I can’t do any work without that extension.

I’m curious if the android version supports this extension?

I agree! My PhD thesis is on this topic [1]. We’ve also done a very limited pilot test on a patient with ALS, with above random chance. Actual results may vary heavily on individual disease progression—the more motor recruitment that’s intact, the better.

[1] https://neuroscience.stanford.edu/research/funded-research/s...

Thanks! That is interesting and well reasoned.

If one wants to change neural computation via a global oscillator, it would seem that a flashing light, auditory tone, or tactile vibration might be a more potent alternative. These stimuli indisputably drive substantial brain wide activity, including many action potentials. I suppose I’m reluctantly team Firing Rate when it comes to mammals—there’s beautiful work on spike ordering in cold blooded animals for example but I’m less aware of computational work where precise spike time matters and not the relative time ie hebbiwn plasticity.

I’m also somewhat influenced by my first hand difficulty when doing in vitro patch clamp to induce sub threshold voltage changes on a neuron with a second probe even just 100um away. Color me skeptical, in part as it’s really hard to do a double-blinded control and the effect sizes seem small.

Cadaver studies have shown that this is almost certainly placebo. Not enough current actually reaches the brain: https://buzsakilab.com/wp/wp-content/uploads/2017/07/Underwo...

We are massively far away from modeling the human brain. First of all, no one can agree what level is necessary to model the brain, and that varies tremendously by scientific question. Personally, my lower limit would be something like the computational package Neuron which models voltages across axon compartments and distribution of ion channels, My upper limit confidence bound is we don’t care about anything subatomic.

At the upper bound: In molecular dynamics, which is used extensively in modern day neuroscience to understand the function of ion channels and GPCRs, a single H100 can model 70ns/day of compute for 1M atoms. There are 8.64e+13 nanoseconds per day. There are ~10^26 atoms in a human brain. Therefore, an upper limit back of envelope is you need fewer than 10e+26 atoms / 10e+9 atoms * 8.64e+13 ns / 70 ns = 1.23e+29 H100 GPUs.

Calculating the lower bound is more difficult, but let’s start by saying you can get away with a fp16 for each synapse. Storing the weights of that model for 100 trillion synapses is 200 Terabytes, and if you figure weight size * 4 or so to do anything useful then this is in spitting distance. Note that this example lower bound is massively less complex than the Neuron model I suggested, as the entire field of neuromodulators, homeostatic mechanisms, glia, and more are thrown out, which are all important for modeling how the brain works under certain computational regimes.

The leaders of SushiSwap have long been running a Kleptocracy. About a year ago, “FrogNation” attempted a takeover of SushiSwap, and as a member of the DAO I did due diligence on the on Dani Sestagelli and Sifu. I found blockchain transactions demonstrating that they had been personally withdrawing >$30M from the treasury of their project WonderlandTime, and posted this evidence to the SushiSwap forums.

I was immediately banned from the Sushi forums and discord. About 30 days later, it was revealed that Sifu was known fraudster Michael Patryn, and the project imploded.

Sadly, the Sushi leadership has continued to suppress any investigative work by DAO members into the actions of leadership. There is undoubtedly a culture of suppressing information, and receipts suggestive of fraud. The SEC is right to investigate.

I feel this struggle although wonder what you recommend instead (besides slack ;)?

Installing home-manager on Ubuntu is practically a blend of NixOS and Ubuntu.