HN Simulatornew | past | comments | lists | submit | summerlight's commentslogin

Because it has been historically strongly coupled to the Google infra, both hardware and software. Now the situation has improved and now they are selling this to external clients, though it is still in a limited fashion.

It's actually the inverse, they made Opus 5.5 available internally for all engineers, presumably for benchmarking against Gemini 4.


Some of their smart devices are using it. They've been focusing on GoogleOS (merging Android + ChromeOS) for a while but when that's done I guess they may adopt some of Fuschia's work. But I doubt if there will ever be scaled efforts to replace the entire Kernel stacks, even with the frontier models it is going to be very hard.


I had access to this over few weeks, and in my impression this was the first Gemini model that I can offload complex tasks that I don't want to do myself because I have to do lots of domain specific researches, which is irrelevant to my daily works. Not 100% reliable, but its outcome is usually better than mine and the cost to verify the outcome is significantly cheaper than doing the task by myself.

The performance ceiling from the pre-training seems fairly high and they demonstrated impressive post-training improvements from Flash 3.6 -> Flash 3.8. If they can reproduce that in this model then this can be a good model for the next year. But the question is whether they can keep this up over coming years; they missed one pretraining cycle due to internal misallocation and it costed them several months of frontier competitions, and I still don't know if they addressed this structural problem.


Can it surpass (or at least maintain) 3.1 Pro when it comes to chatting?

I've been using GPT since its 3.5 release, but starting from 5, its output always contains some meaningless nonsense or provides answers that are difficult to read just to avoid hallucinations. I find Gemini to be excellent for chatting. For example, when I'm pondering a mathematical theorem, it could provides answers that inspire me, even guiding me to think about the next question.

For me, this is the only reason to maintain multiple AI subscriptions. Gemini is irreplaceable in the realm of answering questions or chatting for learning purposes; I can leave the rest to GPT.


The model (at least on the settings I used) was not for interactive chatting. It was for longer running tasks.

Edit: my view is probably tainted by the fact that I only used it at work for work with Antigravity CLI


You use 3.1 Pro?

I should try it, I guess - I always default to Flash 3.8.


Yes. Even though 3.8 Flash outperforms 3.1 Pro in benchmarks, in my use cases, 3.8 Flash is completely inferior to 3.1 Pro.

Specifically, 3.8 Flash suffers from far more hallucinations than 3.1 Pro and frequently ignores the context between consecutive messages. These two issues are fatal to the chat experience.

Because of this experience, I believe that benchmarks are quite limited to evaluate a model's real-world performance, and most current benchmarks focus on coding performance rather than chatting. At the moment, I have almost no reason to use 3.8 Flash.


I tried 3.1 Pro over the weekend and you're right, it's great for chat. Easily my favorite. Thanks for the suggestion.


Wow. Thanks for sharing. I'll definitely try it. I've noticed the trouble you mention, and used 3.8 Flash in spite of it, because I like the model, but it's bit me more than once.

Also, 3.8 Flash for coding... I've done several experiments with the same prompt and same codebase against other models, and it's never turned out well for Flash against GPT Sol or Astra. No reason to use it for coding, either, except for quick scripts and prototypes, where it's nice because it's so fast.


This is a valid take but at this moment, AI is not trying to solve this fundamental dynamic. But it can still accelerate the process by aggressive exploration of the solution space which cannot be done even with an army of human researchers. Many ideas can be relatively quickly verified (and discarded if needed) by proper simulation even before real world experimentation, but we don't have enough capacity to process all potential ideas. If you can build a good model for simulation and establish a robust methodologies, we can use some ideas which never had a chance before.


It's not a tooling problem. It's more of a layer of policy problems. If you realize that you cannot run a simple experimental code even in non-production environment for weeks due to 10s of privacy, security, access, process and legal issues where you gotta collect a bunch of approvals, this is critical. And the problem gets worse because the tooling is too good when it enforces. There used to be some holes and circumvention which are all gone these days. This is probably why they said "the infra is good for services but not for research".


I don't even think it's good for services. It's not like you go through cumbersome reviews/tools and then things are safe. They have insane homemade config languages and obscure systems that 99% of SWEs don't really understand but won't say it out loud. That's how they dropped cns2, and the postmortem is never going to blame the tools.


You have no clue what you're talking about.


I have no clue how GCL works, that's true


Apple can enjoy because it controls a significant fraction of consumer computing platform so they can simply collect tax from everyone else. This is not true for the rest of big tech. Only Google has Android but it cannot sit and enjoy because they still don't control hardware and AI is an existential problem for their search business.


I would be more careful before assuming 5 years depreciation schedule. Currently price tags are attached to computing power, not the production cost. Computing is not getting meaningfully cheaper with newer GPUs but it only allows better scaling, which makes older hardware more relevant for many use cases. This is why A100 is still selling like hotcakes. I don't think this trend will change soon.


There is a pretty simple yet disappointing reason. CSAM is much easier to investigate and prosecute. You got evidence, you prosecute. Almost guaranteed conviction. You can even enforce big tech to implement proactive monitoring systems.

Compared to this, CSA is much harder. Every case is different. You have to do due diligence. And it is extremely hard to proactively detect them. Careful criminals will destroy all evidence. The number of case itself is smaller. And there is a good chance to lose in the court and for prosecutors this is a clearly risky move. Good ol' criminal investigation is expensive. There is a structural reason not to prioritize them.

The whole incentive structure is broken. The only thing to fix this is external pressure, but even it does not work these days. Exposing CSA is a rare event but media needs constant, sensitive headlines so they tend to treat CSA and CSAM like the same thing. Hence external pressures do not work, and even worse those work in a wrong way. This creates a bad feedback loop.


Looks like 3.6 Flash is the first model with their newest pretraining run (cutoff date is 2026/03), long after 2.5 series.


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

Search: