Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Our eighth generation TPUs: two chips for the agentic era

blog.google

Post reply on HN
Print view View original

231–240 of 240 posts

Re: Our eighth generation TPUs: two chips for the agentic era

#231

Post by gnunez »

VectorLock wrote: post #75
akersten wrote: post #43

Earlier quoted context omitted.

I'd go long Google too if using Gemini CLI felt anything close to the experience I get with Codex or Claude. They might have great hardware but it's worthless if their flagship coding agent gets stuck in loops trying to find the end of turn token.

I use Claude Code all day and use Gemini CLI for personal projects and I don't see the huge gap that other people seem to talk about a lot. Truthfully there are parts of Gemini CLI I like better than Claude Code.

I agree. I like using Antigravity for some of my frontend work, and I find it does a better job than Claude Code - Opus 4.6. I’ve also found the Gemini Flash models to be good at legal defense research—I use them to help New Yorkers fight parking tickets (https://nyceasyparking.com). That said, the Claude models are still amazing at agentic work.

Re: Our eighth generation TPUs: two chips for the agentic era

#232

Post by jauntywundrkind »

nsteel wrote: post #130
jauntywundrkind wrote: post #124

I'm surprised the interconnect per system is so slow? 6x 200Gb feels barely competitive. Same as last year. Trainium3 and Maia 200 are 2.5 and 2.8Tb/s vs this 1.2Tb/s. Maia is 6 stacks of HBMe3, so ratio of mem:interconnect bandwidth is really falling behind here. Notably Maia is also, like TPU, high radix.

Isn't it 6x 200Gb octals? An octal being 8x 200Gb lanes. So 9.6Tbps?

Thanks. Yeah uhhh the table here says 19.2Tb/s scale up per chip??? Uhhhhhhhh. This math is still not mathing for me. But that makes much much more sense.

Google just wildly ahead of literally everyone here.

Re: Our eighth generation TPUs: two chips for the agentic era

#233

Post by svnt »

WarmWash wrote: post #229
svnt wrote: post #220

Earlier quoted context omitted.

Good luck all-in. But seriously what are you doing that this works? I guess if you are writing pop culture articles this might work. For anything where the output has consequences I can’t imagine finding success like this.

The latent space knowledge that the models have is stronger than the inference agent going out and trying to find information to integrate into context. If you ask why the sky is blue, the model already has the answer. It's corrosive to your conversation to pull a bunch of unknown sources into context so the model can appease your "feels right" request. If you don't trust the answer, your brain is still way way bette…

I use it claude and gemini all the time and they get more advanced theory, motivation, and history wrong all the time.

If you aren’t seeing the errors it is because you are in some really mainstream conversations or because you don’t know what they are saying that is wrong.

This is trivial to demonstrate to yourself for any nontrivial project. A single academic question is easy to get the right answer for. That is not the dominant AI use case for most product people or engineers.

Re: Our eighth generation TPUs: two chips for the agentic era

#234

Post by deaux »

awongh wrote: post #175
orbital-decay wrote: post #162

Earlier quoted context omitted.

Regarding Anthropic, they used to make best multilingual and generalist models, it's their policy thing, not a capability issue. Claude 3 was best at this, including dead and low-resource languages. Neither modern Claude nor Gemini are remotely close to what Claude 3 was capable of (e.g. zero-shot writing styles). Anthropic basically reversed their "character training" policy and started optimizing their models for c…

The benchmarks don’t seem to say that language ability has gotten worse?

There are no real benchmarks of how "natural/idiomatic" output is in a multitude of languages.

"Multilingual benchmarks" are usually something like "How good is it at a multiple choice exam like the SAT in language X". This is a completely unrelated metric.

Re: Our eighth generation TPUs: two chips for the agentic era

#235

Post by deaux »

orbital-decay wrote: post #162
deaux wrote: post #151

Earlier quoted context omitted.

> If I had to guess the pro and flash variants are 5x to 10x smaller than opus and gpt-5 class models. I really doubt it, especially Pro. If anything I wouldn't be surprised if their hardware lets them run bigger models more cheaply and quickly than the others. Pro is probably smaller than GPT 5.4 and Opus 4.6 (looks like 4.7 decreased in size), but 5x seems way too much. IMO Gemini 3 Pro is the most "intelligent" in…

Regarding Anthropic, they used to make best multilingual and generalist models, it's their policy thing, not a capability issue. Claude 3 was best at this, including dead and low-resource languages. Neither modern Claude nor Gemini are remotely close to what Claude 3 was capable of (e.g. zero-shot writing styles). Anthropic basically reversed their "character training" policy and started optimizing their models for c…

I've never ever had Gemini over the API switch languages in translation tasks and that's across more than 10 language pairs and 6 figures of calls, across both short and long outputs. Maybe your languages are even lower resource ones, though we do include Central Asian languages.

The Chinese models are very prone to it, they love to mix them up.

I've seen it in chat, but IMO that's more of a system prompt/harness issue.

I'll admit I don't remember Claude 3, the oldest data I have seems to be 3.5. And at that time Gemini 1.5 Pro did a much better job across all of our language pairs, it wasn't close.

Re: Our eighth generation TPUs: two chips for the agentic era

#236

Post by deaux »

blueblisters wrote: post #217
deaux wrote: post #151

Earlier quoted context omitted.

> If I had to guess the pro and flash variants are 5x to 10x smaller than opus and gpt-5 class models. I really doubt it, especially Pro. If anything I wouldn't be surprised if their hardware lets them run bigger models more cheaply and quickly than the others. Pro is probably smaller than GPT 5.4 and Opus 4.6 (looks like 4.7 decreased in size), but 5x seems way too much. IMO Gemini 3 Pro is the most "intelligent" in…

3/3.1 Pro appears to have knowledge about eccentric topics with no obvious sources that often turns out to be right. It does hallucinate a lot though, and is the most affected by context rot in multi-turn conversations

Agreed on both, especially hallucination. That's what makes their chat app even worse, it's very opaque about web search and sources, so you can't tell whether it's a hallucination.

Re: Our eighth generation TPUs: two chips for the agentic era

#237

Post by nsteel »

jauntywundrkind wrote: post #232
nsteel wrote: post #130

Earlier quoted context omitted.

Isn't it 6x 200Gb octals? An octal being 8x 200Gb lanes. So 9.6Tbps?

Thanks. Yeah uhhh the table here says 19.2Tb/s scale up per chip??? Uhhhhhhhh. This math is still not mathing for me. But that makes much much more sense. Google just wildly ahead of literally everyone here.

19.2Tb is 2 x 9.6Tbps. Some (most) companies count Tx and Rx separately despite it making no sense in the context of serdes lanes. Stupid marketing in my opinion.

I don't think they are meaningfully ahead, it's more to do with what's available at the time. 200/224G is only just coming available this year. The others will have the same in their next product announcements.

Re: Our eighth generation TPUs: two chips for the agentic era

#238

Post by awongh »

deaux wrote: post #234
awongh wrote: post #175

Earlier quoted context omitted.

The benchmarks don’t seem to say that language ability has gotten worse?

There are no real benchmarks of how "natural/idiomatic" output is in a multitude of languages. "Multilingual benchmarks" are usually something like "How good is it at a multiple choice exam like the SAT in language X". This is a completely unrelated metric.

then there should be such a benchmark :)

Re: Our eighth generation TPUs: two chips for the agentic era

#239

Post by foobazgt »

chaoz_ wrote: post #226
delbronski wrote: post #181

Earlier quoted context omitted.

I think so. From my experience Claude/codex tooling really excels at vibe coding the whole thing. You give it a folder and just say: now make it do this. And you don’t really care for the code. Junie tooling excels when you are more involved. Like, look in these two files, add this specific functionality, in this specific way. Junie is usually a lot faster and to the point. Very simple tooling , it just works for thi…

It surprises me how JetBrains managed to lose such a great market opportunity. I don't think they ever going to be able to re-claim large chunk of developers who are now fine with thin VSCode-like + Terminal for non-JVM languages. Perfect example of how large corp with research capacity failed to navigate their product changes.

My most enjoyable and productive experiences with AI so far have looked more like pair-programming than agent-based vibe coding. That is to say, I care about the details, and I want to read, understand, edit, and curate the codebase. I find that if I'm not limiting AI to relatively small enhancements per request-review cycle (100 or so LOC), then when things inevitably go off the rails, I'm in a deep hole that takes a long time to climb out of.

I haven't tried out Junie yet, but the concept seems pretty compelling to me. I want a good IDE for the language I'm using, and I'd like an AI that's well integrated and trained on delegating to it for algorithmic/deterministic transforms (e.g. IDE-driven refactorings).

Re: Our eighth generation TPUs: two chips for the agentic era

#240

Post by jauntywundrkind »

nsteel wrote: post #237
jauntywundrkind wrote: post #232

Earlier quoted context omitted.

Thanks. Yeah uhhh the table here says 19.2Tb/s scale up per chip??? Uhhhhhhhh. This math is still not mathing for me. But that makes much much more sense. Google just wildly ahead of literally everyone here.

19.2Tb is 2 x 9.6Tbps. Some (most) companies count Tx and Rx separately despite it making no sense in the context of serdes lanes. Stupid marketing in my opinion. I don't think they are meaningfully ahead, it's more to do with what's available at the time. 200/224G is only just coming available this year. The others will have the same in their next product announcements.

Thanks again. Very PS, rad to see you putting commits into mopidy still. I haven't touched mpd like systems in a decade but they have a fond place in my heart!!
Post reply on HN