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.
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.
Author here. So what! I am not talking about promising directions for AGI, I am talking about having computer systems that we can have confidence in. Sure, AGI if it ever happens will look more like emergence than compositionality, and I'm sure it won't feel a need to explain to us fallible humans why its decisions are correct. In the meantime, I'd like computer systems to be manageable, reliable, transparent, and accountable.
100% agree. When we explicitly segment and compose AI components, we are removing the ability for them to learn their own pathways between the components. We've been proven time and time again the bitter lesson[1]: that throwing a ton of data and compute at a model yields better results than what we could come up with.
That said, we can still isolate and modify parts of a network, and combine models trained for different tasks. But you need to break things down into components after the fact, instead of beforehand, in order to get the benefits of learning via scale of data + compute.
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
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.
Correct me if I’m wrong - but the novel thing is not that it’s possible for ultrasound to pass through the skull, but that it’s possible for it to pass through the skull and back in a way that an image can be reconstructed.
> OpenWater's Transcranial Focused Ultrasound Platform. open-LIFU is an ultrasound platform designed to help researchers transmit focused ultrasound beams into subject’s brains, so that those researchers can learn more about how different types of ultrasound beams interact with the neurons in the brain. Unlike other focused ultrasound systems which are aimed only by their placement on the head, open-LIFU uses an array to precisely steer the ultrasound focus to the target location, while its wearable small size allows transmission through the forehead into a precise spot location in the brain even while the patient is moving.
FWIU NIRS is sufficient for most nontherepeautic diagnostics though. (Non-optogenetically, infrared light stimulates neuronal growth, and blue and green lights inhibit neuronal growth)
A commercial medical ultrasound imaging device in doppler mode can pick up and map onto the image plane some of the vessels in the brain through the skulls. But mostly just through the temporal bones(where the skulls is like 1-2mm thick). (The commercial machines run doppler on lower frequency than imaging signal so you get no s tructural image this way, only the color doppler map(unless you find a place in the skull where an emissary vein passes through the bone table where the image signal can ride through))
Through the temporal bone of most people you can catch some sparse doppler signals with average hospital gear.
The fontanelles enable good ultrasound imaging on an entirely different level. A highres greyscale image vs a few sparse blobs of doppler from major vessels.
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
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.
> "directly print pure copper, which has historically been very difficult"
SLM [1] has been able to 3D print Copper with precision down to the size of a mechanical pencil's lead for a long time already. In what way is ECAM better? Is it more precision + no need to handle powder + no need for laser source and containment - ECAM being slower, or am I missing some crucial feature?
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.
SLM machines typically use an Argon gas chamber. DED machines use an Argon gas shield.
> It’s more typical to print copper alloys than pure copper.
In the context of modern SLM, it depends on your definition of "pure" and "alloy". During the process, a bit of resin to is mixed into the powder and heat treated in a final step to get to 99.9% pure copper.
edit: Just fixed up my knowledge. Indeed alloys are typically used (99% copper with things like Chrome added on depending on use-case), tough the pure copper can be used with higher laser power.
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.
Sorry I wasn't talking about density but the copper content of a powder which is printable. Googling a bit I found this presentation from 2022 showing that a density of 99.5% for pure copper is possible although at half the productivity of a copper alloy https://www.coppercouncil.org/wp-content/uploads/2022/02/TS2...
The copper use-case is what kick-ed off an industry-wide race towards offering blue laser as an option. There is more than just wavelength that goes into printing good copper results, but that is a major factor.
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.
reply