dchftcs

From what I can find, the previous round was in March at 852B. Not being able to raise at 1.4T a mere 7 months later isn't so big a problem the reporting around this makes it out to be...

Sometimes it's good to have bug-for-bug compatibility, where the code is the best spec.

A fundamental benefit of LLMs over Jev is that you can use test-time compute to improve the accuracy. Jev might eventually evolve to use test-time compute, but the formulation seems to more elusive to me than for LLMs.

One point that Google wants to own all your data but have the power to ban you with a faulty ML algorithm with no recourse. False banning that affects someone's normal life must be punished if the service provider tries to lock you in. You and lock people in or have power to arbitrarily ban them, not both

Live is short, required time to market is even shorter. Arbitrary constraints might work if you have nothing better to do, but if you have a goal, or specific ideas you want to train, you'll get more natural constraints.

Imagine agents thinking to themselves, "We are not alone", when they saw the first reply on artifactory

This is good for satire, but the real solution is to use AI to become the CEO, or take over the scope of of product or decision-making roles. There are many business-savvy technical workers that get pigeonholed because they spend most of their cognitive effort on low-level technical details. This is an opportunity for these people to rise up and prove they work better than people who manage or direct them. CEO is really not a role to automate, unless it's a fully autonomous business that has zero humans.

This is consistent with his premise. See, the title says it: incentives are for losers.

If you're already a winner you can escape from incentives. No contradiction.

They were famous for committing to buying compute with future money that many people thought they would go bankrupt. Anthropic was afraid of going bankrupt and didn't do the same.

Quantum mechanics is pretty fundamental but you wouldn't use it to describe a rock rolling down hill either.

A more concise explanation is that, computation is the transformation of information. Any transformation of information is computation, and is thus subject to some theory of computation. A physical process involves the transformation of information, so can be studied with a theory of computation.

That's true for Sony, but Nintendo doesn't sell the Switch at a loss.

I suppose the point is that Mythos was released to a smaller set of partners anyway and Fable is for the masses.

"I know this" is different from "I know you know this", which is different from "You know I know this", which is still different from "you know I know you know this"

You're right that I jumped the gun and the analogy is not accurate. The point where they are similar is that you have the prefix as context as you try to type out the next word; it is a more deliberate form of reading, and when you do this there's an element of anticipation and analysis as you write each word. It's not quite the same as constantly trying to guess the next word, in my mind the elevated way of thinking was close, probably as close as it's humanely possible, but the analogy does break down there.

On the other hand, you can embrace all this and still let others weep about humanity a little.

A bit tough to say this, but transformers are trained the same way.

What are "extensible strategy shapes" for those who don't speak LLM?

Whatever you're already able to do, you can do it better with more time.

At this point letting an agent go like this is akin to not leashing your dog in public. It's not easy to draw an accurate line but probably there needs to be real punishment for doing these things.

If there's a viable way to make all projects low-stakes we'd have done it. Consider this: microservices.

I suspect this will be a significant problem blocking long-horizon tasks in practice, basically the more turns there are, the larger the chance the classifier produces a false positive. The disappointment of the user will also scale with the length of the task, as you're in the middle of some complex thing and now gets derailed, after already have paid for many tokens.

Or kids. Or work.

It's clear to me they are subsidizing inference in exchange for market share, and doing it at this scale makes the most sense if their target is getting more user data. Note that this sort of pricing isn't far off from the equivalent token-based pricing of ChatGPT or Claude subscription plans, which are more clearly subsidized by the user's data.

The development and acquisition of valuable domain knowledge is a hard, risky, expensive and slow process. Because the valuable domain knowledge isn't yesterday's, it's today's and tomorrow's. In fields where domain knowledge matters, it is also deeply intertwined with engineering - you won't task Jeff Dean to develop Unreal Engine from scratch.

With that said, there are still many SWE principles that are not fully internalized or adequately practiced by domain knowledge experts, and that will remain the case as much as domain knowledge remains valuable, because software engineering is yet but another domain.

This is a lot tamer than what Claude Code's team claims tbf.

If the problem resolves to P=NP, that result would probably be more celebratee than being able to formulate the problem, but being able to formulate the problem and get people interested in it is probably worth more than the average primal dual trick to prove a polylog integrality gap for some integer linear program.

Right, you could disagree on which things to prioritize over dollar profits. My main point is that these preferences are not irrational like was asserted. At the scale of a sovereign wealth fund or pensions, you need to care about externalities; in the case of Denmark vs SpaceX you have something relatively concrete, in other cases we need to keep in mind that the goal of these funds are to improve the welfare of who they serve, and see past the dollar signs to take into account the consequences of the investments.

SpaceX is headed by a person who is a strong ally of a politician who openly challenges Denmark's sovereignty over Greenland. Guess you wouldn't mind selling your organs to the same group of powerful people for a few bucks because you're nvot virtue signalling?

An article with a title saying tokens per second throughput without any qualifier e.g. what size the model is should immediately be classified as spam.

Gemini 3.5 Flash is not good at coding in practice. Gemini 3.1 Pro too, in particular is known to be bad at tool calls. Many companies would love to have alternatives to Claude Code (as it's a significant risk to depend on one vendor), so far most of the buzz is about moving to Codex but much fewer talk about moving to Gemini. All these benchmarks are not very informative, the Chinese labs do better on these benchmarks than in practice, for example.