root_axis

It's no more a perversion of writing than the act of using an LLM to write in the first place.

The "untestable" meme is just a thought terminating cliche. It's as if you believe no researchers ever considered the brilliant insight that the framework is difficult to test. History is littered with plenty of examples of a novel discovery, understanding, or technology yielding new opportunities for experimental design.

It's also a glib dismissal of the unexpected mathematical elegance of string theory that makes it so compelling. Worse, it completely ignores the material contributions to applied physics that string theory has made possible which is obviously worth "spending your life's work on" - as if people need to justify the value of their life's work to disinterested onlookers lest they be demeaned.

> My idea is mixing deterministic nature of existing tooling (non-LLM tooling) with non-deterministic nature of LLMs.

This is the fundamental idea behind every LLM harness.

TypeScript's type system is extremely expressive while still allowing you to retain the flexibility of a scripting language. v8 and JSC also have decades of performance tuning across basically every consumer device.

"Gambling on a bailout" doesn't sound like an appealing investment prospect.

To be fair, a plurality of voters care more about the culture war than the economy as well.

> which is now on much shakier footing

Is it though? As far as I can tell they continue to maintain total domination of web search, and LLMs do not replace search engines, they work on top of them.

Ok, now do memory capacity and bandwidth - the things that actually constraint local LLMs.

It sure did. Even Al Capone went to prison.

Send you to jail.

Physical reality is not software or math, LLMs can't design "equipment to manufacture autonomously".

Sure, I'm just saying, even for "mundane" computing tasks where the hardware is cheap, if you're optimizing for cost, renting hardware is always the better choice.

> A lot of people spent a LOT more money in every election for many candidates

And in this case the topic of discussion is how one specific individual spends their money - so Biden and Democrats and whatever other axe you have to grind is totally irrelevant.

Well, if you only ever needed a $20 subscription's worth of tokens it would never make any sense to buy hardware since even at 2018 prices you could buy 20 years of subscriptions.

I agree, but I'd say there's obviously some verbal intelligence involved in regurgitating arbitrary knowledge on demand in a semi-convincing way. It's absolutely a stochastic parrot, but parrots have some intelligence.

I'm pretty skeptical of the LLM hype, but I'd say it's fair to describe LLMs as displaying some form verbal intelligence.

"Knowing the codebase" seems like an antithetical philosophy to LLM driven engineering orgs.

> which can be read as implying that the rewrite hasn't been used in production.

Well it hasn't. "Production" for a language runtime means being generally available for arbitrary use. The engineers that built the runtime using it to release one closed source binary to the public is, at best, an extremely narrow beta test.

You're right. I misread your comment as attached to a different thread.

> Can't we just go back to node + npm + vitest + vite?

You're kind of stating the reason for the buzz there: one runtime that does everything you need rather than cobbling together a dozen different tools.

Not that there's a problem cobbling together different tools made to do a job, each tool has a purpose and solves a problem, but having all problems solved out of the box is very convenient.

Bun does offer an additional major advantage, i.e. the bundler has a runtime API, so the same process that serves your assets can also bundle them in memory without having to coordinate an external bundler writing static files to disk.

The fact that you're likening a sparing few narrowly targeted tarrifs to capriciously applied illegal global tarrifs suggests you're not quite looking at things fairly.

Whether or not it's illegal is typically determined a few months after the act in question.

Since 2024, SCOTUS has sided with Trump about 90% of the time, so most likely it's legal.

Still, even if it's later determined to be illegal, since the law presumes the president has legal immunity to all crimes committed in a presidential capacity, there's no legal risk, so the damage can be done long before the court makes it illegal. For example, Trump's global tarrifs were determined to be illegal a year after the fact, long after the damage had been done.

Have you been paying attention the last two years? With a few exceptions, he has unilateral authority.

LLMs heavily rely on traditional search engines behind the scenes.

No he wasn't, stop spreading lies.

Because the model can output data in a manner optimized for training a new model, including outputs that were post-trained like RLHF and RLVR.

Same experience I have with bun in general. As an idea, it's the absolute best js runtime by a very large margin, but in terms of stability it's terrible and segfaults 20x as much as node (I base that number on newrelic telemetry)

> Phones have gigabytes today.

Phones have zero HBM today.

> There are famous laws of growth that put terabytes at just a few years away

I assume you're referring to Moore's law here, but if you are, it doesn't really apply to HBM in the same way, especially in a smartphone form factor where LPDDR is the only practical option due to heat and energy constraints, and a variety of architectural complexities specific to HBM that make a TB of it in a smart phone something far beyond what we can hope for in any timeline we can project today.

Not really. Models that can run on a $250 card may be able to produce a decent word doc sometimes, but the fidelity of such models is so poor that it's no longer very attractive for general use on those types of tasks.

> But how many years do you guess? I personally do not think it will take even 10 years for the situation to be commonplace.

IMO it won't be possible for the foreseeable future. There's essentially zero possibility that phones will gain the hardware capacity to run today's Kimi, so the only other alternative is to squeeze the power of today's Kimi into something that can fit on a smartphone, which also seems fairly unlikely considering the current rate of progress.