vehas

@sun_hanchi @jukan05 I believe this too but do you have evidence for that ?

@jukan05 No, when 950 comes online, the price will drop even further

Does this mean DeepSeek has already brought a Huawei Ascend 950 cluster online?

banger

Should we expect reset every time we have new competitor model released? That would be cool

welcome to new genimi time capsule edition

@ClementDelangue No body outside unknown enterprise have access to it though

I remember when people were saying "It's useless to open-source big models because nobody will be able to run them fast"....

@_simonsmith I think that the point. put the risk downstream to token buyer, suck capital from hyper scaler And package in such a way that buyer think they are at low risk by always have options to sell that token later

Honest question: What's to stop a company from buying compute at a discounted rate and then reselling it in some way? How much of a wrapper is needed to avoid getting cut off by OpenAI? Because if anyone can just do this, it strikes me that the biggest buyers will be investors who create a token futures market.

@nivi it's the model that getting really good. 5.5 is the beast.

Memory in ChatGPT has gotten INSANELY good and nobody is talking about it. It shocks me daily.

the fact that new @grok product name grok build 3 mo after "claude cowork" is just embarrassing. at lease have confident in yourself and try to compete in knowledge work space not just software-bro.

@ivanburazin frontier lab will come up with one, how can I any one compete with that ? + unlimited token to build, + 50x token discount on sub to lure you in ?

A cloud built specifically for agents is coming. It will have: - web search - sandboxes - databases - storage + all the primitives (purpose-built for agent access patterns). Hopefully, it will be the answer to the unpredictable demand spikes brought on by agents. It'll look more like Stripe than AWS (a consumption API, not a general-purpose cloud) OpenAI wants to build it. Vercel wants to build it. We want to build it. It's the next addition to the stack and one of the biggest prizes in infra rn.

@myfirstmilpod @amasad Day 1 1M arr is 2,700 usd on the first day Isn’t that much impressive isn’t it.

Replit's founder made $3M in 2 days shipping a half-ready product. Amjad Masad's team wasn't ready to launch Replit Agent. He said: "I don't care if it's semi-broken. If it works 50% of the time, it'll wow the world." For the first time ever, an AI agent could write code, debug it, create a database, and deploy to the cloud. End-to-end. He shot a video on his iPhone. Posted it. Andrej Karpathy quote-tweeted: "I feel-the-AGI moment." OpenAI and Anthropic researchers reached out: "We didn't know our models could do that." Day 1: $1M ARR Day 2: $2M ARR In 48 hours, he made more than 8 years of trying. Amjad: "Product-market fit feels like stepping on a landmine. We pivoted for years. That moment was the landmine." If it doesn't feel like a landmine, you haven't found it yet. Full episode: youtube.com/watch?v=ddSucX… @amasad @thesamparr @ShaanVP

the question is not "Will AI replace my job", but "how can we develop expertise with AI" what does it's look like ?

Wen gpt-5.5-spark ? need it.

@drummatick another W for OpenAI compaction then.

A major reason why LLMs performance degrade as context increases is we simply DO NOT have that long multi-turn conversation in the data distribution. Instruction tuning dataset rarely has data which has multi turn chat spanning millions of tokens while being coherent. This lead to poor instrunction following and hallucinations. So it's better to compact the context much much before you hit 1M token context. From 200k onwards the returns feel diminishing

@DaveShapi > Because if our tax dollars fund the training runs that produce ASI, it belongs to all of us. Full stop. Drug industry have other idea, and they already win

I lowkey believe that AI will become so expensive (per Epoch AI) that eventually the ONLY way to train new frontier models will be through corporate consortiums and maybe even governments pooling resources. Which may mean that superintelligence is automatically democratic. Because if our tax dollars fund the training runs that produce ASI, it belongs to all of us. Full stop. And, even if it is privately funded (imagine Google, Meta, Microsoft, Amazon, IBM, Oracle, etc) all pooling their resources together to train the models, they will probably choose some form of open source so that they all equally benefit from it. I mean look at Nvidia, they are starting to train gigantic open source models because they don't give a fuck whose model is running so long as it's running on their GPUs. I know I've been somewhat more openly cynical about power structures and profit motives lately, but I see this as a real possibility. I mean, how else are we going to train frontier models once the price tag is $1T? That, or we just stop training frontier models for a while and wait for the hardware to catch up.

If it is that good why does they just release it right now ?

@JawardSesay_ Does this mean Anthropic don’t have any better RL tool up their sleeve?

It took GRPO to get Claude to reason in natural language:)

codex team ? in codex desktop app are there any change that I can pay for 2x or 4x for only image gen? it's really slow right now, and I'll gen 10x more image if it's 4x faster

@tocophonic @adrianvalbuenai @claudeai @SpaceX Some loser discounts maybe, but the winner ? nah.

@amvehas @adrianvalbuenai @claudeai @SpaceX After the bubble has bursted, 147B seem adequate.

We’ve agreed to a partnership with @SpaceX that will substantially increase our compute capacity. This, along with our other recent compute deals, means that we’ve been able to increase our usage limits for Claude Code and the Claude API.

@adrianvalbuenai @claudeai @SpaceX Ai industry pump 700B/yr and you think 147B can buy winner ? Really ?

@claudeai @SpaceX Meanwhile, Apple is sitting on a $147 billion pile of cash waiting to buy the winner of the war.

It’s bound to happen, and it’s total Elon L