swordsmith

Seems very oriented toward model architecture and inference engineering. Maybe add some more on model training flow, distillation, data generation, SFT and RL techniques?

Seems like he thinks RLVR == learning from binary reward for the whole chain, completely discounting techniques to provide denser rewards like process reward supervision?

> No one wants 125 million thoroughly non-Sinicized Japanese inside the country; that would be seen as an endless headache, not a prize.

I don't think what you claim the people want matters (if even true). Look at Tibet and Xinjiang

lol you can go ask Thomas Rearden himself if Ctrl-labs "invented" gesture recognition from sEMG.

The technology was "invented" by CTRL-Labs like how OpenAI "invented" transformer-based language models.

Virtual keyboard is completely doable, but too slow.

He's scribbling with his finger.

Typing can also work, but handwriting is simply faster and easier to decode.

sEMG signals correlate with *muscle* activation. When your fingers move, the actuators are the muscles in your forearm, and the tendons relay the force on the joint. Placing the band higher up on the forearm would actually give you better signals, but a wrist placement is much more socially acceptable.

One device at a time!

Everyone focusing on consumer prices. But tariffs also function to incentivize domestic reindustrialization, which has huge national security implications. You see this clearly in the venture space as increased investment interest in hardtech and manufacturing. It's great that the federal government is looking long term again.

Came to the comments to see this. Stamping has all sorts of problems with alignment, stamp resolution, contact quality, etc, it's not clear whether this so called "simplicity" still holds after scaling up the resolution even more.

Thanks for the reference, I suspected it would be EMG as well. Especially in the video you can see how the patient modulates his eyebrow, facial muscles, and mouth. The vestigial muscle movements can be decoded to speech with the help of LLM much more easily. Actually, if form factor was not a concern, this can be done even more easily with other sensors as well.

I use Foundry for work. It makes data ingestion, cleaning, quality check and automation easy. After all the data is ingested, running analysis/RAG on them become extremely easy.

Basically, it's end-to-end data engineering and analytics. And the more a company uses/invests into the platform, the more benefit and locked-in they are.

Hard agree. This type of app looks awesome in theory, but also makes me want to really organize my thoughts, make sure the tags are not too redundant, and think about how to merge with my previous notes (a lot of which is on paper). Then the perceived work involved for using such an app just for taking spontaneous notes become way too much.

> Sometimes I think that part of this site's audience only considers art from a consumer-product perspective

Imagine you are an artist. You as a patron connect and talk about a piece of art, you walk away happy and feeling connected. He does not sell any art because they are indistinguishable from what the AI can produce with the cost of electricity. Now if you are willing to pay for "bonding over human experience", GPT can probably be trained to do that as well. Generative AI eliminates the livelihood of almost all artists eventually.

> I've been seeing content on Tiktok where even 100 miles away they're seeing small dead fish which is apparently according to the folks making the videos a massive red flag.

It's interesting that you mention that you saw this on TikTok. In the last two days, the Ohio train derailment has become one of the hottest topics in ALL of Chinese social media (Weibo, tiktok, Little red book, etc), 10 days after its happening.

I was confused when my Chinese friends started asking me about this event, which I saw on the news the day it happened.

Why the huge lag? Oh yeah, because of the Chinese "balloons" -- nationalist citizens are mad that CCP has taken a "soft" stance in the face of US shooting down Chinese balloons. Chinese official statements are along the lines of "how dare you shoot down my balloons, they are just passing through on accident", then stating that American balloons have entered Chinese airspace previously and that's unacceptable.

This entire thing is a mess, so here comes the typical media manipulation to focus both Chinese and non-Chinese negative attention on America, even though American media has been giving this event no less attention than something like this usually receives.

That may be the case in the past, but the factors that enabled the Chinese manufacturing (low to medium range) dominance are largely declining: abundant growing labor force, stable world trading and geopolitical environment.

> I seem to recall Xerox Parc was designed to foster serendipitous meetings between departments.

This has been a common argument about returning to office, especially for R&D jobs. But in my experience in corp R&D, people are so busy in meetings/running between conference rooms, there isn't even enough time to get their actual work done, let alone do free-form discussion. ICs end up using the after-work hours to do most of their work.

WFH actually solves that -- people can stay longer for meetings near the end of the day because they are already home, and don't need to beat the traffic.

> Humans weren't meant to live like this.

I agree. It would do those people good to get a life outside of work...

> I’m open to being proved ignorant here. Can you think of some examples where tech was obviously ahead of its time and not accepted?

Google glass as an AR device

So with these generative models running rampant, what's to even motivate aspiring artists to develop and hone their craft, if their years of work can be copied so easily?

Maybe it doesn't practically matter, because some art style generative model can be developed and feed into the diffusion model so it can generate art of new styles.

I actually LoL'd reading this. You can buy packs of 10 small notebooks (each the thickness of 10 of these cards) for $15 on Amazon, they work just as well if not better. A notebook is much easier to carry than a stack of cards with a wooden holder. It can be stood up against the monitor, water bottle, or even by itself.

The bubble symbols...spoiler it's very easy to mark each item with "*", "+", "-", and any other arbitrary symbol with a pen.

Great marketing on this product though, I'm sure many would buy it as a status symbol and make themselves feel better for having such an elegant productivity system.

These other countries don't have calls to defund the police for one

Does this provide real time "coaching"/suggestions during the game for League? I remember there used to be this application called "bwcoach" for Starcraft Broodwar, that would yell at the player if he's not building enough workers, not spending enough resources, not scouting, etc.

The capabilities of your League app seem to be easily accessible via multiple other apps (porofessor, op.gg, League client itself), and during pick/ban, suggesting champions/runes to use based purely on community stats may actually be detrimental (the player may be much worse at a suggested counterpick, than if they just played their best champion).

I think if the coach client can provide suggestions such as the following, it would provide a lot of utility for minimal effort: 1) reminder to check mini-map 2) after a kill, suggest backing vs. pushing wave to tower vs. roaming 3) itemization suggestion, which has a more objectively optimal choice based on team comp/itemization than champion selection.

+1 on this. Matlab toolboxes are much more well tested, has thorough documentation and work much better out of the box, especially for controls and signal processing. Working with scipy can be a pain sometimes with incorrect or unstable results.

> Here's another from 2010 with a similar result under similar conditions: https://pure.ulster.ac.uk/ws/files/11410334/cecotti_tnsre.pd...

Within the abstract:

> The average accuracy and information transfer rate are 92.25% and 37.62 bits per minute, which is translated in the speller with an average speed of 5.51 letters per minute.

5.51 letters per minute, not words. The work you cited is not comparable to the UCSF work at all.

It seems you are measuring performance in terms of bit-rate (i.e. based on how many symbols per minute), which makes sense when you are using a cursor-based speller.

This approach is not a correct measurement of bitrate with this speech-motor decoder, however, as words are being decoded based on the syllables contained within it. The decoding model is trained to recognize 50 specific combinations of syllables, and the total number of unique single syllable phonemes is about 44.

From your top level comment:

> a nurse a laminated piece of A4 paper and a patient who can blink get about 15 characters

How is this a relevant comparison? 15 characters per minute is much less than the 15 WORDS per minute performance this work demonstrates

> I spent a solid 6 months intimately learning the pros and cons of various existing BCI systems, as well as the exact methods researchers use to make their technology look like a breakthrough when it isn't.

Then you should know this is a huge deal.

> Specifically, here, they've created a system that can decode one of 50 symbols

Previous motor and speech neuroprosthetic systems focused on pointing (controlling cursor) and more recently, handwriting (https://www.nature.com/articles/s41586-021-03506-2). This work goes gives communication rate similar to the handwriting work, but decoding speech from the motor cortex has been much less understood than that of simple motor movements such as cursor position and velocity control.

Even more impressive, the test subject is not even a native English speaker.

> They've also cherry-picked the patient that responded the best.

Bravo-1 is the only subject they had.

> When you peel back the hype, this is about the same performance we've been getting since '08 for an invasive, subdural, non-penetrative array.

This is completely, utterly false. See (https://stacks.stanford.edu/file/druid:jx921pv3255/Technical...) for survey of performance of typing BCI.

You seem to love to cite your background as a BCI PhD dropout. I should also point out that I have a completed PhD in invasive neuroprosthetics and still work in the field (not that matters when anyone can look up the sources and judge for themselves).

Sure, a word can evoke all sorts of meaning, other brain process, etc. The decoding principle leveraged here is exactly as I indicated. It's decoding motor activation associated with imagining speaking a word, from the speech-motor cortex.

The study you quoted uses similar principles.

Unless you have years of experience at IC design companies, where would you even get the experience for IC design and implantation without a PhD?

That's not talent flight. He was likely pushed out. But Elon's companies are a grinder though, one of the cofounders Vannessa Toulosa is now at FB.

> true consumer BCI does require an implant.

Good luck convincing millions of people getting a brain implant.