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

It's obviously not an option to just build more.

Success is clearly doom.

Failure is the path to prosperity.


CXMT is going to make enough DRAM soon enough to more than make up for whatever was lost converting DRAM fabs to HBM.

Less familiar with the NAND market.


Is that so? Everything that I have read indicates that CXMT capacity will take a year or so to come online and it's not much capacity (at first) neither.

Consumer electronics product design cycles run 2-3 years out, so if it takes a year to come online that's being reflected in pricing and purchase decisions made now.

Why wouldn't CMXT end up selling to Chinese Cloud companies? They are building datacenters too.

It's a global market. China's demand being satisfied by CMXT means they're buying less from the other manufacturers, and the rest of us will face less competition for Micron, SK Hynix, etc.

The same/opposite reason gas prices went up in the US after the conflict turned hot with Iran, even though the US is normally a net exporter.

Of course it's not quite that simple, but it'll do for a decent first order approximation.


And why would CXMT miss the hype train others are riding?

Suppose they have the technology to make DDR4 and DDR5 but not HBM. What are they going to do? Make nothing until they can develop the technology to make HBM, or make as much DDR4 and DDR5 as they can in the meantime?

Also, the incumbents are converting existing fabs to HBM because it's faster than building new fabs. CXMT has to build new fabs anyway, and then it's not a matter of which is "more" profitable since they can do both in parallel. X + Y > X even if X > Y.


Can you trust Chinese memory? Will you try it out if it's cheaper?

It's an easy time to grab market share.


Can you trust Chinese everything else? Lots of people do

Because HBM is an integrated product. Nvidia etc would actively have to partner with the Chinese HBM suppliers. So it is a much safer bet for CXMT to sell non-integrated DRAM chips. It's a commodity market after all.

I suspect they may want to get away from the size class due to the fact that - at the pace they are launching new models - and the similarity of model sizes between Neon/Carbon, Opus/Fable, Sol/Astra - the best model is going to be the most recently released one, not the biggest one.

I'm not sure it's much better than 4.1 Flash vs 4 Pro -> if 4.1 Flash is better temporarily when Pro is typically significantly better, people are going to be confused.

I'm not really sure how you get around this problem easily.

It's not going to age well when the next Flash-lite model launches, but maybe they will call that `Gemini 4 -lite` - who knows...


Interesting. You're claiming that the Opus/Sol/Neon class and Fable/Astra/Carbon class models are similar sizes to each other? That would be surprising to me.

I've found that while Opus 5.5 is often pretty nice for practical work, Fable (even 5.0) is still significantly smarter on hard tasks and makes far fewer "dumb mistakes"


A Microsoft Azure page briefly claimed Sol runs with less looped transformer passes; and so you could speculate that Sol and Astra might share weights; but have different “loop” factors.

That would make a lot of sense from an inference optimisation perspective. The same prefill and decode machines can hold one set of weights, and serve both Astra and Sol requests.

At (multi-)terabyte-scale model weights, and when you have to offer not just the latest like Fable 5.2 but still serve 5.1, 5.0, etc; heck I think Opus 4.6 is still around; you really care about overall compute utilisation.

Idle GPUs are wasted GPUs; and even over the fastest interlinks you’re not hotswapping multi terabytes of weights.

With MoEs, there’s probably stuff you can play around with varying the number of active experts; or other clever optimisations.

Maybe Sonnet 5.5 could be just Opus 5.5 with a different router and less active params? Who knows; I’m just theorycrafting now :)


> A Microsoft Azure page briefly claimed Sol runs with less looped transformer passes;

This doesn't really make sense given how the models released.

OpenAI released Astra long before latent-space looping was a thing at the frontier space.

Then they released Sol 6, which was fast, clearly not looped. Then they release Sol 6.1 which was nearly as good as Astra, but way cheaper, and WAY slower -> this would follow it just being a looped version of Sol 6.

It doesn't seem likely that Astra 6.2 is just a looped Astra 6... It's too fast, and they'd go bankrupt if they're charging the same price... It may be a further looped version of Sol than Sol 6.2 - which would explain why it's not as much better as we would've hoped.

It may also explain why their release is not going as well as they hoped. It's hard to predict how your first looped models are going to fare.

If you look at how long it took Google to recover from their failed 3.5 Pro, it would be surprising if GPT recovered from their rumored failed run on Astra in under a month. That would give them reason to release a looped Sol and call it Astra.


Looped transformers have been known for years now. I don't think it's something new.

They are JUST arriving to frontier models.

Even in the open weight Chinese models as small as DeepSeek V4.1 Flash - nothing is looped.


I think they're responding to the claim Astra can't be looped, there's clear and convincing research, as well as public comms, that it is. It's unclear why Astra can't be looped because looping is new, but, 6.1 Sol released a week later can be.

Astra was the first shipping looper. Unconvincing evidence would be reporting x OpenAI's quick reaction to it via a blog post from their chief research (it broke what's considered a safety taboo). Convincing evidence would be research from ex. Neel Nanda on serial reasoning, recent update at https://x.com/EpochAIResearch/status/2122762411287748696?a=J...

I heard rumours that Sol 6.1 was cut down Astra because Sol 6.0 was received badly. But I thought that was the exception (and a result of Open AI's actual Sol-class model being uncompetitive) rather than the rule.

That would line up - on release and for the few following weeks Sol 6.1 felt like Astra but maybe more task oriented, recently it’s been a noticeable downgrade in the level of specification needed to successfully set off.

Fable & Opus, Astra & Sol, and Carbon are all thought to be >1T param models - AFAIK, most people think Opus & Sol are ~5T param models and Astra, Fable are 10T+.

I think the rumor-mill consensus is that Google's models are significantly smaller while attempting to swing at the same level - but the ratios AFAIK are thought to be around the same - i.e. Neon is somewhere around half as big as Carbon, big enough that a newer checkpoint can make it temporarily a little bit better.


So many figures and ratios pulled from a single ass

There's only a limited amount of hardware for training and serving models at scale, and which form an upper bound for what model size/inference can support.

A GB300 rack has 72x288GB=20TB VRAM. Do the math and 10T parameter models are just about the max size you can easily train and serve (at 1M context) without doing a lot of custom hardware development or limiting your scale.


The Bitter Lesson - Richard Sutton (2019)

https://news.ycombinator.com/item?id=19393432


> Gemini 4 -lite

Surprised if it’s not going to be 4 electron/proton.


This model feels like it might actually be as good as Opus 5.5.

This is probably the set of benchmarks most people care about: https://www.vals.ai/comparison?m=claude-fable-5-2,cobalt-c-h...

Overall, performs just under Fable (above Astra or Sol 6.2 or Carbon - slightly ahead of Opus 5.5).

Latency is by far the best for the level of quality.

It's about 2x more expensive than Sol 6.2, but significantly better and faster. You get what you pay for.


I am a Claude Ultra subscriber that's very happy with Opus 5.5.

If I wanted to realistically try out Neon in some of my workflows without committing to a paid subscription before seeing if it can actually work as well as Opus, do I have any option, or do I just need to take the gamble?


I was thinking of giving Gemini 4 Neon a try as a code reviewer instead of GPT-6.2 Sol, and found this in their FAQ:

"Using third-party software, tools, or services to access Antigravity is a violation of our Terms of Service"

and also: https://antigravity.google/docs/faq?can-i-use-antigravity-cl...

I'd like to keep my Gmail account, so Google can keep their model and will not get my business.


You can't use subsidized tokens in 3rd party apps. Although they just suspend your Gemini account now, not your google account.

That being said, you can do whatever you want with API tokens.


You can with Anthropic and OpenAI, using 'claude -p' or 'codex exec'.

Paying for Gemini at API rate is therefore out of the question for me.


you can use agy in batch and/or JSON mode - it's what I do for few of my projects. You can also use "agy acp" to get ACP server now (used to be a separate binary)

The requirement is that you use antigravity harness to access subscription-based accounts, not that you can't wrap it


Their official wording seems to be that it's not supported:

> Wrapping agy inside third-party coding agents or applications, or automating account rotation, is not supported under consumer plans.

At least they no longer ban the whole Google account, so I think interpreting that as a prohibition only for abuse would be reasonable.


That violates the ToS for both, and they absolutely will ban you (there's tons of examples of people complaining about their cool tool to get Claude Code shimmed over an API not working/getting them banned, same with OAI).

They won't ban you for that.

I never had any issue using Claude for code review in Codex and vice versa.

I remember that half a year ago, Anthropic planned a change to make claude -p draw from a separate usage pool at API pricing, but after feedback that change was scrapped.

What people were banned for is using the subscription authentication in a third party tool, using many accounts, or reselling.

Invoking the official CLI from another agent for legitimate personal use during development should not be a problem.


I think that page is out of date. You can use it via T3 Code now and Google has said it's okay. It's via their official ACP-compatible endpoint that's now provided by the antigravity-cli.

You can install Codex in Antigravity its kinda fun running both.

You don’t need to use the models through antigravity. You can get them through ai studio (or whatever Google calls it) and pay api pricing to use it via whatever harness you want.

They say it's going to be available for Pro users - so you don't need a $200 account (Ultra). But, I don't think you're going to be able to try it unless you get at least a $20 account.

IIUC, you also don't get to try Opus unless you have a subscription, right?


Reminder that Google trains on your data if you're using a consumer sub whereas Anthropic and OpenAI claim they don't if you've got all relevant boxes turned off. I know everyone stopped caring about this at the start of the year and believes the latter two don't honor it anyway, but it's good to know.

How do you find Fable 5.2? I tried it for a few days for coding, and I'm back on Opus 5.5, and even Deepseek 4.1 Flash for simple stuff. It was just going off on tangents way more for my workflow.

Glad I'm not the only one. I found Fable 5.2 to be a substantial step back vs Opus 5.5. It would happily burn a weeks worth of Ultra usage in a day to output an absolute slop pile of wrong assumptions and bad decisions. It just sucks for agentic use or even review work. Initial design back and forth is the only place I could get it to be useful. I'm starting to wonder if all the amazing examples anthropic gives for it only work if it has effectively an unlimited token budget so it can spin it's wheels for however long it takes to eventually get a decent result.

You want a Google AI Studio API key.

You can use Neon on OpenRouter.

openrouter or api directly? great question to use your claude sub to figure out easily :)

You can always pay API pricing if you don’t want a subscription?

OpenRouter

0% chance this is better than any frontier OpenAI model.

How can you be so confident about a model that was released today? I just dont get the lack of epistemic humility here

>how can one presume the sun will come up tomorrow merely because it has come up every day prior? have some epistemic humility! i am a very smart rationalist.

Edit: I tested the model and it found several mistakes Astra 6.2 made. It’s possible…


I think that OAI likely has the strongest internal model, but I don't think they can figure out how to ensure alignment on it

I also have the hottest wife but no one can see her because people wouldn't be able to constrain their lust. I need to hide her to maintain safety.

However hot she is, she isn't solving millennium problems.

There's thousands of benchmarks, and nobody can agree on what the best models are if you run all of the benchmarks which costs tens of thousands of dollars for frontier models.

Then, somebody runs a small subset of benchmarks - with not enough data to have any statistical significance - and makes further claims.

How are we supposed to derive value from this?


Although people disagree with the best models, the worst model is obvious.

TL;DR - from the little I could gather, Sol 6.2 looks particularly impressive.

But... is it just me, or does this product launch page look like AI Slop?

https://www.vals.ai/comparison?m=claude-fable-5-2,gpt-6.2-as... seems like a better comparison than what I could get from their page.

Astra is ~20% slower, ~40% cheaper, and a bit less powerful than Fable. Sol remains way cheaper than anything remotely in the same league, and it's speed to complete tasks seems to have improved considerably.

Scrolling through it has the same hypnotizing effect of trying to read anything written by LLMs.

It seems like things that should make sense, then I try to look at the charts and a lot of them are comically unreadable.

I tried interacting with some of the examples, and the Lamp Builder - for example - I couldn't even figure out what it's trying to show or what I'm supposed to do - or why I should be impressed.


I wouldn't go as far as to call this slop but the lamp builder thing was super weird and unclear for me too! Even clicking the other tab cause my browser to scroll down a lot which I found very jarring.

> If Nvidia's next CEO is smart, they'd buy up a bunch of these companies that "burn" weights to silicon,

Models are getting significantly better every 4-8 weeks with post training now.

I don't think there'd be much demand for this.

~80% of the total cost of inference is chip depreciation. If inference is 10x more efficient, that only drops your running cost by about 5% absolute tops, and that's assuming the chips cost the same and not more for this. That's also assuming the chips have the same lifecycle, which they wouldn't.

No one in the world is going to want Gemini 3.5 Flash at a 5% discount 1 year in the future. You've taken a 3-year chip amo, turned it into a 1-year, and that's 80% of the cost.

Hard pass.


> Nothing seems functional nowadays

It also wasn't functional before-a-days [=


Now we can make even less functional code much more quickly! All it costs us is some dignity, poor people's electricity rates going up and massive, massive amounts of cash set on fire

Such a deal.

Now we have the added benefit of nobody understands it.

> Now we have the added benefit of nobody understands it.

You should peak inside some old COBOL codebases running some of the most important software in the world [=


> I've yet to see anyone with expertise in medical research express optimism that AI will cure major diseases like cancer in the next several years.

Depending on "several years" -> it is almost by definition impossible. The trial periods are too long.

A lot of major breakthroughs have happened recently and are already in the pipeline, looking promising.

There's a lot of progress on this stuff already. Compared to 10-20-30 years ago, a lot of diseases already look "cured", even if they aren't completely.

Near immediate death sentences have become 5-10+ years of life expectations, etc.

I don't see any reason why AI wouldn't help speed things up in the future.

But we won't see any of that for "several years" regardless.


The trials argument is a good one, but if a good cure is found, trials can be sped up. A lot of medication today has an impact that's hard to measure, so large and long trials are necessary. If a cure is so good that improvements can be seen immediately, I can imagine that they are greenlighted faster, especially if the alternative for patients is death, since what side effect could be worse than death.

> This should be fine for caching.

When launching a swarm of smaller models, they can theoretically share cache amongst themselves, but not with the larger model (that already has it cached). They have different architectures.

Further, it wouldn't really make sense for them to share the same cache anyway! The swarm, that is, they'd be working on independent things...


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

Search: