chrishare

I agree that hindsight is doing work here, but DeepSeek R1 from Jan 2025 seemed to heavily leverage distillation, and 18 months is an eternity in this climate.

Super interesting. I am always wondering where the bridge between symbolic programs and inline connectionist approaches could be. LLMs can call traditional programs "in process" as tools, but the other way usually just looks an external, expensive call. Maybe this fits in for a certain class of program. I will definitely add some support to my little collection of browser based AI tools https://andergrove.com/tools/ai/

For super simple programs, yes, this doesn't make sense. For very complicated programs, the generated code would begin to look like spaghetti so fine that it begins to look like weights.

1/ Agreed, better naming convention and model layout 2/ It isn't, there would be many more comparison benchmark results if it were, but also - theatrics may be marketing 3/ Disagree that cheaper models don't have a place 4/ Do they need to keep up? 5/ It's boring until something you own or run gets compromised, I guess, but even then - this is preview of things to come (biosecurity, etc)

Rough but true. This has PMF but not at this price.

Well, it probably is an advantage over the same product written in say C++, all other variables held the same - just not a big one.

Selfishly, I hope this doesn't reduce to 0 the amount of time he spends doing educational content, which seems like a particular strength of his. I presume this means Eureka Labs is not releasing any product or course.

LLMs are definitely intelligent - just not general like humans, and very very jagged (succeedingand failing in head-scratching ways).

Yeah, but atleast the dog is going to eat your documents only, and not crap on your rug

Does the lowercase convey authenticity or lack of care?

Good doesn't wash away bad

Link?

It's this I believe: https://www.w3.org/TR/vc-data-model-2.0/

This is certainly no fatal for OpenAI, but there is some irony that Altman and Musk are both struggling.

All of this is true and credit assignment is hard, but the brutal competition between Chinese firms, especially in manufacturing, differentiates them from and advances them over economies in the west. It makes investment hard as profits are competed away, which is blasphemy in Thiel's worldview, but is excellent for consumers both local and global.

It's not illegal, you just get shot is all

I think the Tailwind case is more complicated than this, but yes - I think it's reasonable to want to contribute something to the common good but fear that the value will disproportionally go to AI companies and shareholders.

Not at all - it's legal, but it doesn't garner goodwill either.

When training for muscle size atleast, but not strength. Presumably there are increased injury risks overall when lifting heavy (based on a brief search).

Well, many companies are still mispriced, but stock markets today look a lot more like voting machines than weighing machines - so it takes more time to be proven right (or wrong).

Have you tried not being poor?

Sergey Levine, one of the co-founders, sat for an excellent Dwarkesh podcast episode this year, which I thoroughly recommend.

He's referring to humanity, I believe

Yeah, more or less. Being in the application space as well as the inference space hedges a variety of risks, that inference margins will squeeze, that competition will continue to increase, etc etc.

Their contribution to opensouurce and open research is far behind other organisations like Meta and Mistral, as welcome as their recent model release is. Former security researchers like Jan Leike commonly cite a lack of organisational focus on security as a reason for leaving.

Not sure specifically what the commenter is referring to re: scammy, but things like the Scarlett Johansson / Her voice imitation and copyright infringement come to mind for me.

True, but there are many reasons besides. Meta and Anthropic attract less criticism for a reason.

What was it?

Most would be search engine agreements I presume, which is still proportional to the user counts.

Well, hopefully you can find some suppliers you can settle down with and live happily ever after

Nitpick - it's the ML system that is sampling from model predictions that has a temperature parameter, not the model itself. Temperature and even model aside, there are other sources of randomness like the underlying hardware that can cause the havoc you describe.