vessenes

Also I note from your comments here you don’t plan on local inference - that seems like a huge self own, provided the model is less than like 2b params. I’m not saying people won’t in practice prefer your cloud version, but you have this huge uphill battle ahead of you on privacy which can be wiped at the stroke of a pen with “yep it will run on your Mac.”

You can get both markets (ease of use and technical) by allowing it. Also i would bet that you gain almost no product security locking off local inference - the hardware is the hard bit. Once someone has your hardware in hand it will not take them long to reverse and build their own processing layer; you’ll need to fight those people on patents and product quality.

Heck for a v1 just allowing local sensor data would probably be enough for the tech enthusiasts. But please don’t lock this off from the get go - it makes the sale harder and I don’t think it gives you much in return.

The Gemma task is harder: add features. Qwen was: fix this looping behavior

It’s a good reminder of one of the most unique features of alphago - it was like a seven layer or so neural net that could predict out a hundred moves. I’m not sure we took this kind of compression super seriously or at least I would say perhaps it remains under explored. You could think of what you did as sort of the same thing as skipping the Monte Carlo search phase for a go engine — a process compression.

Good points - I’m still shook

Yes - “opus 5 helped train 5.5” is extremely cool but not shocking. Hand editing weights without tooling to change outcomes is shocking.

A single forward pass here Bette 10mm+ PPO rounds. You are hard to impress.

Super cool. For a cool voice change demo we'll allow a gb200 over low latency wifi :) elevenlabs has deals with a number of professional actors and voice actors for licensed uses, seems like you could probably find an ethical and cool co-brand opportunity in there.

Have fun building this! It's worth doing.

This is insane.

Astra can, without tooling, determine what games small networks are trained on, can edit the weights of networks directly to improve outcomes, and can hand edit a small qwen model to improve its math skills by like 5x.

I'm shook. If you'd asked me this morning whether LLMs have native understanding of their weights that is less than human <-> neuron understanding, the same as human or more than human, I would have said "same to less" depending on definitions.

No human can go edit a brain to become better at math. This is just very, very surprising to me and extremely interesting. I would guess this upsets quite a number of interp theories and philosophical takes on what's going on with these networks.

I mean we're getting pretty close here -- you could have these + your AirPods in transparency mode up and you're not far off. Also Claude won't currently ice you out for going dark, unlike Jane.

Just preordered -- glasses request: there's no way to say it nicely, I have a big nose. With a high bridge. Many glasses are really tough to get properly placed; may I request you not give up on the pads w/metal adjusters that you have in the product photos? Don't let a designer take them away from you, us folks not one standard deviation from 'normal' noses will thank you.

Second, how tough is this signal processing? Will it be done on-glasses, or through an app? Third, what's the equivalent of a wake word for these? I'm reminded of the Apple Vision Pro radar to see if your fingers are clicking. I could imagine a lot of modalities where I want to just start talking and have things happen, and a lot where I don't want the glasses to pick it up. Fourth, I think a cool voice change demo would be rad -- not sure what latency is like, but if you could get it end to end under like 250ms you could probably have someone exaggerate the mouth movements and pipe out audio of them talking like someone else. Which would be another really cool demo.

Super cool. You're going to be working 80 hour weeks for the next year, I'm excited to see what you put out!

Ooh, cool. I'd like me some luminous concrete please. I like these order of magnitude changes in common technology -- in this case, way smaller minimum size and more efficient per square nanometer -- the use cases language is a good reminder that the innovation part ("Look! This is cool!") often (usually?) precedes the most compelling use cases.

I imagine this is like 20 years away from hitting consumer design, but I bet it'll be used for some cool stuff.

It's clickbait. They don't do anything like prove the platonic representation hypothesis.

They do seem to have some really interesting training efficiency improvements, though. The paper is at least worth a read.

That's also almost certainly not a cat.

Thank you. I missed the gvs5h tie in.

This working journal is so much slop it’s basically impossible to read. Perhaps the author could write a short human written and readable summary so we can evaluate if this is interesting or not.

Nice. I like that different print runs will have different procedurally generated flowers; nice touch.

There are lots of measures in place at these exchanges - the rub is that cold storage is by design slow. To service withdrawal requests in a timely fashion a percent is kept “warm” or hot and used to process withdrawals. Usually this is where the leaks happen.

Tesla

Do laser printers include tracking dots? I thought that was a color printer thing.

This could be another point informing your prior as to how much benchmaxxing and distilling of closed models is being performed by the open weights publishers as well.

I don't understand this conspiracy/suspicion vibe around AI benchmarks -- the best benchmarks get maxed quickly in today's world, especially if the datasets are open. We're in a situation where the AA's of the world show 0-1 point difference between models, so by definition they are NOT good discriminators. These tests NEED improvement.

Do they work with frontier labs? Probably. They should work with them at least a little so that we can get an independent read on the model quality. Do they inject new tests that specifically favor one particular lab? Clearly no, as different labs have moved up over time. Do they inject new tests that help to provide some separation on scoring? God, I hope so.

Nota bene, another change they announced was dramatically increasing the amount of held back test data; this would be a simple explanation for seeing open weights models, all of which are known to distill from frontier labs and benchmarks, drop in the rankings.

Deepmind has over a decade of DNA of not releasing - it's a significant cultural difference.

I mean, it will pay off for society one way or the other — we’ll have the datacenters and (hopefully) the weights at the end of this.

That’s an amazing amount of revenue and spend, though. Wild.

They've always been realllly slow to roll out stuff outside the org when it comes to AI. Most googlers themselves likely won't have access to this model internally.

These are right at the edge of desirable for me right now -- but it's hard to pull the trigger. I'm imagining an M7 Pro with 480 GB/s and therefore an M7 Max 850-900 GB/s. And significantly better GPU speeds as well. If I knew I could get the M7 Max Q2 of next year, especially at 192GB of RAM, I'd wait.

But then you have to compare to buying a GB200, sticking it in a closet and sharing it out in a workgroup and having a lower specced laptop. Decisions..

Interesting idea -- when everything's on a prediction market, all speech can be a securities fraud!

The article is confusing - at one point it says the truck left the road, hit a bunch of stuff and caught on fire. At another it says they don't know if it was a battery pack fire that spread. I'll be interested to read the final safety reports and see if this was just an unlucky crash or something really bad happened with all that lithium.

I'd really love a followup post where you find the highest scoring pizza possible -- feels like you must min-max now, yes?

Oooh this is cool. Create gothic-esque designs in stone construction using CNC. I like the apology right at the top: "medieval builders were always on the forefront of technology". Word.