seizethecheese

To me, "a probably $200-300B company" sounds absurd. You're describing one of the possible outcomes of a massive range that goes from $0, where they get over their skis on financial commitments to well beyond $2T, where their 10x YoY growth continues for a few more years. A $200-300B outcome is possible, but certainly not probable. (And in fact, for us HN commenters, the most probable outcome should be assumed to be roughly what the market will bear.)

This feels like an ex post explanation for 20th century history.

For example, the Great Depression. Hoover was a literal textbook author, but is widely blamed.

The only real professor was Wilson, but he had similar failings to what people in this thread are worried about.

Okay so assuming engineer compensation of $200k, you’re suggesting something like a thousand-fold productivity increase on this task in terms of cost and 26x or more in terms of time?

> Stocks going up has been so normal for so long that it doesn't get much press. Any significant drop is way more newsworthy.

It's true that stocks tend to go up, but it's also true that large daily moves are pretty close to 50/50. It's brownian motion with a trend. The drops really do get more press than the spikes, and I don't think it's because they're less common, it's because negative news catches more attention.

I run startup so I set these standards in a small way. I aggressively promote AI use, but do not allow AI writing in Slack. Two reasons:

1. The writing is terrible to read and wastes time.

2. The act of writing ensures understanding. If no understanding is needed, then this doesn’t need to be in Slack, Claude could be just doing it.

The major buyers of inference are not government. The major funders are not government. And the industry isn’t predicated on some sort of government backstop (like perhaps banking).

What’s your proposed mechanism here? Maybe data center buildout risk, but that seems mostly local policy?

Somewhat worse, it’s actually if you serve any model at all, you need a license.

Since the reaction here seems to be universally negative, I’m curious what people see as the benefit of having several gig worker delivery platforms.

As a thought experiment, we currently have a duopoly for ride share. Would we be better off if Uber were broken up into two platforms?

Obviously having one platform would be bad, since there would be no price competition, but what’s the optimal number? Surely, too many platforms would be bad, as the concentration of delivery workers would be terrible on each app, making the whole system inefficient (setting aside whether one things the system should exist at all).

Interesting perspective.

They do serve the same market for gig workers that don’t want to be ride share drivers, though, right? I wonder what their market share is for that.

I mean, running for office is essentially a full time job right?

It seems totally reasonable to both have a law that protects a citizen’s right to run for office without losing their job, but it also seems reasonable to require some sort of coordination with an employer.

(Knowing nothing about this, it is seems like it should work roughly like an extended jury duty.)

it's opencode

Wow, cool, thanks!

> anthropic only prompt caching

Anthropic has custom prompt cache requirements, but we've verified all providers are working well.

> Blinded model.api.id = "contestant" bypasses gemini.txt and serves Claude's prompt (default.txt)

This is wrong. We're serving default.txt which is opencode's default prompt. We are not serving a custom prompt for any model, including Gemini.

> Corrupted/dropped Gemini thought_signature (reasoning_details) in tool loops

We blind the reasoning details in the frontend because this would leak information. Some models summarize thinking while others provide raw thinking. I'll check if we're corrupting the thought signature that's actually sent to the backend.

So far this sort of looks like an example of Gemini being bad. Was this Argon?

Only 3.8 Flash which also didn't do particularly well.

It's possible, but we are getting real prompts from people on real projects, and the instructions about the battle are pretty minimal.

Oh, sure, but isn't claude code still generating this by default? We're just inheriting opencode behavior.

Okay, this is interesting. We are hardcoding the effort (to improve data concentration), and Argon is hardcoded to "high".

Yeah, this is a new project so help me out here. I mean that the harness is running locally. I'm trying to distinguish this from arena.ai webdev arena, where they run the agent in a cloud container.

So far in our data, there's no difference between Wonder and Sol 6.1. This is with wide error bars, as we're just getting started. Note this is Wonder, not Full as it's not on openrouter yet.

But yes, Opus is a tier above Wonder and Argon with very high confidence.

This would make sense. In our local coding agent battle system, Argon is tanking, and this is running on opencode, importing agents.md and claude.md.

Yeah, we are doing something different and actually running parallel local coding agents head to head, something Arena doesn't do.

Our data is all software engineering run via local coding agents, but yeah your specific use case could be one where Argon is stronger. (Or your personal preference.)

You're right Arena.ai has far more data. Our data is actual local coding agent head-to-head battles, so I think our quality is far higher.

Their agents leaderboard is literally just randomizing the agent and looking at metadata like tool call errors. In my opinion, it's poor quality data in high volume.

Their web dev leaderboard is in cloud containers, most people don't run agents like that.

I ran a statistical test on our data. Deepseek vs. Argon. Posterior probability DeepSeek 4.1 Flash is stronger: 99.1%.

We're working on a battle system for local coding agents [1] and Argon is really bad. Literally the worst model in our initial ranking, and Deepseek Flash, Muse Wonder, Grok are all in there.

Top is Opus 5.5, followed by Fable.

[1] https://github.com/bottomless/agent-duel

I've done a lot of PCBA sourcing from Shenzhen so I have domain knowledge here.

The main latency for this is typically sourcing. The partners I've worked with will get PCBA done within a day or two once they have parts.

The claims here about a breakthrough in setup time also don't really fit with my experience. What is this, putting a new stencil in, adding different components to a hopper? I've toured PCBA factories and it's not like their lines are typically down for set-up. I think this takes like 30min?

Also, it doesn't look like they substantiate their 2x speed claims? They "benchmarked" this, but with what suppliers? Their claims would make sense if they found random suppliers, not exactly specialized in rapid prototype style PCBA, so they're not comparing like to like.

Don't get me wrong, I really want this to happen too, but the desire for this to happen seems to be making the community gullible. I'd love the founders or someone with circuithub to link to real benchmark evidence and explain why their changeover procedure can produce several days of difference.

I think the premise is interesting (that there are no breakthroughs in cryptography) but does not support the conclusion (that therefore three letter agencies are suppressing this).

I think the premise does support the conclusion that there's some mechanism suppressing release, though.

Valid criticism, but I recall the study showing cab drivers have the lowest incidence of dementia among all professions, and I thing perhaps there could be something about activating the visual system in 3d space that is uniquely good.

My dad was a newspaper reporter in the early 2000s, so there were several dinner table conversations about this dilemma.

The buyouts were taken by superstars and low performers. The superstars moved on to other things, like working in other forms of media.

Dad didn't take several rounds of buyouts and was fine.

After suffering a stroke, he was "put out to pasture", which for a journalist means he was sent to cover city hall meetings in some suburb. Eventually, he retired with some much smaller payout than what the last buyout would have given him. The sense in the family was that the collapse of the industry denied him a dignified last decade of work. (When the industry was flush with cash, older journalists were sort of like "elder statesmen" in the newsroom. Post-internet, no longer.)

Today’s prices are a ceiling for today’s capabilities, but not at all a ceiling in general. One could imagine dramatically higher spend for dramatically higher performance.

This is how pretty much all S-1s look, it’s because if they fail to disclose a risk that comes to pass, they are in hot water.

Expertise is the raw material here, not the same as the end product.

Its always a question of why businesses don't just eat the next level, downstream.

Why doesn't NVIDIA just make the best frontier model?

Why doesn't TSMC just make the best chip design?

Why doesn't ASML just make a fab?

The real answer here is that OpenAI and Anthropic are in the middle of an intense race, and the payoff of having the best model (or one of the two best) is so large that it makes no sense to focus downstream.