Hacker Newsnew | past | comments | ask | show | jobs | submitlogin

First impression: Third-party benchmarks or gtfo. Personally, I've never heard of either of these companies before. We're just supposed to take their word that they've matched the best models on the market?

Sakana describes their model as a "Orchestration Model." Does that mean that it's actually a bunch of different models glued together?



Is it actually that hard to make good models or is it just about the amount of resources you have to do training? (This is an actual question, I really don't know.) I'm sure it's not trivial but does it really take world class secret knowledge to build off of the known existing techniques? I feel like there's tons of low hanging fruit still to explore, and time and resources are the limiting factor.


The gap between grok and Gemini to Claude and chatgpt suggests that yes it is that hard.


All of the 11 grok co-founders alongside Elon quit: https://techcrunch.com/2026/03/28/elon-musks-last-co-founder... so that will have hampered grok.

Like Zuckerberg, top talent may not work with a polarising character if they disagree with his behaviour. Space focused talent don't have many choices aside from SpaceX but ai companies are a plenty and a top AI person can pick and choose.


The fact that you need top talent also suggests that it is indeed that hard


I dunno, if most of the top of a company quits it's extremely disruptive even if everyone else in the company is competent.


I suspect that Grok has been ironically lobotomized by pressures to correct its political views.

Similarly, I could imagine the Gemini folks working in a significantly more complex corporate climate, with different parts of Google pushing for different capability focuses. They are only lagging behind less than a year, so it isn't too large of a gap yet.

That said, the fact that Anthropic is currently the top dog suggests that talent and execution is incredibly important. A year ago none of my normie friends new them, and when i suggested using Claude looked at me like when I recommend Linux.


That shouldn’t affect Grok’ coding ability. How often are people discussing politics with Claude code? Writing decent code is just hard and it’s not just Grok.


It affects their ability to hire and retain talent.


If training a good model requires talent then that’s the answer to the question this thread is trying to answer: is training a good model actually that hard?


Talent to do.. what? This could mean a lot of things.

Navigating astronomically huge fundamentally not so hard but still really tangly and hairy projects requiring both excellent short- and long-term vision in an overheated domain with angry people and lots of money is a skill all of its own.


Talent to train high quality LLMs, especially coding LLMs.


Why would these be independent?


More specifically, political lobotomy shouldn’t affect coding ability.


You’d be quite surprised, I think. Fine tuning a model on one axis can have drastic impacts on another that as a human we would expect to be completely unrelated.


I have never seen anyone argue that this cannot be overcome with more high quality RLVR data.

The practical reality is that the Chinese and American models might have very different politics. But the most relevant factor in model performance is the quality and volume of training data, not ideology of the base model. Unless you are suggesting something very particular about the way Grok was neutered.


It's all a bunch of weights isn't it? Why wouldn't fiddling with some parts of the weights have cascading effects?


Yet empirically it does


Not true, aggressive post training makes models notably dumber.


> That shouldn’t affect Grok’ coding ability.

If you are spending all your time having to re-train because the boss doesn;t like the output, it will hamper coding


> A year ago none of my normie friends new them, and when i suggested using Claude looked at me like when I recommend Linux.

Isn’t that still the case? Normies haven’t even heard about Claude, in my experience.


In my experience it has improved a bit, but 90% of my normies still have no idea. (It was 100% before)


Not hard to be a fast follower. Lots of companies are ~6-9 months behind. Reaching the actual bleeding edge is much harder.


>Is it actually that hard to make good models

Didn't take DeepSeek long. Or XAI to launch grok.

If they have a top team and the money then appears to be a matter of a year or two? And one startup mentioned is Japanese not Chinese so they won't be banned from buying US tech.


Their release post was on HN recently. The comments seemed to think that it was similar to OpenRouter, not an actual model.


My impression is that the answer is yes, that it purports to dispense the glue on-the-fly in some kind of dynamic way rather than being some kind of new model-amalgam.

See also contemporaneous reaction at:

https://news.ycombinator.com/item?id=48624782 (6 days ago, 244 points, 133 comments)


Also sakana has misrepresented their findings previously i think to remember [1]

1.https://www.reddit.com/r/singularity/comments/1iwbwgu/sakana...


Did Anthropic give you third-party benchmarks? Is that what you said to them? Yes, they're important, but the attitude is wrong.


Anthropic always publishes 3p benchmarks every time they announce a new model


No, stop right there. Anything published by Anthropic implicitly is not third party. For it to be third party, the third party has to be the one publishing it.


When you're announcing a new model, typically, nobody else has benchmarked it yet, because it hasn't been released yet. You can still run 3p benchmarks on it and publish those results. If other parties later run the same benchmarks independently, and find major discrepancies, that would be a scandal.


And even if they didn't, they have a track record. Even if we did have benchmarks in this case I would still wait until people got there hands on it and formed a more holistic opinion.


Fudging benchmarks is a cheap way to get attention. If the model is really that good, it will have plenty of attention soon enough.


Yeah, what happened to that scam startup that alleged to have made a model context window breakthrough a few weeks ago?




Consider applying for YC's Fall 2026 batch! Applications are open till July 27.

Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: