pls seed

@DeDgemaskerde @magnushambleton You are, I discovered I can only get a 3.5 beer from the store.

@magnushambleton Europoor!

Worth keeping in mind for the muricans that think we are a socialist hellhole that will confiscate your money!

@Presidentlin oh, come on, all the problems that where there before, none of them are gone now, he just changed his opinion. can happen. when there is a lot of money/influence/ego on line i tend to not believe it.

@YYYYOOOO77 He isn't shilling

What a dark path we have taken. Lisan hater of OAI is being scolded by their CEO to say mean things. Popeamodei, why hast thou forsaken us? Where art thou, O Claude?"

@joshmanders @thdxr @yiannis__p everybody literally complained about github doing. and i dont buy the to train open source models. then make the dataset public from day 1. otherwise its just non binding promise.

@thdxr @yiannis__p Yes and that's kind of the point. The opposite is you're preying on people who are forgetting to opt out. Telemetry should be always opt-in, if that becomes an un-meaningful way to lower costs, then find a better way to lower costs.

opencode go is currently zero data retention however we can increase limits and make it all more sustainable if we collect data to train future open source models you can opt out of this - is that something you'd be ok with?

@thdxr @yiannis__p a come on, you know that its not right and you still want to do it. why did you ask then.

@yiannis__p people always say this but you can understand that it won't be meaningful enough to actually increase limits the opt in is paying $10 for $60 of inference

@tushant_suneja @leonabboud @nikitabier this, you see this, the crypto bots are gone now you have AI bro bots

@leonabboud we ran into similar issues with local hosting and that's why we built Snowy AI, a secure cloud interface that makes it easy to use these tools without the hassle of setup and variable costs

Anthropic banning OpenClaw got me to downgrading my $200/m subscription and I'm now back on the Plus plan for OpenAI. I still think it was a massive fumble loosing AI power users to competition. No one likes pay as you go usage, people would rather have a fixed plan they're on. Just like no one would like paying per day to go to the gym. Even though most would likely save money paying per day than they would paying for the month. If OpenClaw was costing Anthropic a lot, why couldn't they create a new plan called "AI agents" that was slightly more expensive but was a fixed monthly rate. Most, including myself would have been willing to pay a bit more for a plan specifically designed for AI agents.

@initjean where is codex bro ?

idk, but dax are you okay?

@0xSero lol we are inventing IDE backwards.

Warp + Droid is almost perfect. Needs a browser, I need browsers everywhere, please Warp. Still, can't believe they open sourced it. Very top dawg of them.

@SamuelWayne0 @haider1 i find the 3.6 27b with reasoning disabled better, on a 5090 i get like 1.8 ttft and 119tps with 155 context size. this thing rips. thats just vllm with a couple of tweaks and sakamakismile/Qwen3.6-27B-Text-NVFP4-MTP

@haider1 What are your Concrete use cases? & how many TPS are you getting?

google gemma-4 31b is such an underrated beast model it's probably the first consumer hardware-sized model that is genuinely usable for simple conversations, not just specific tasks to put that into context, i prefer it over the free-tier chatgpt model and you should also

@burkov I used it in opencode, works ok, not great not terrible. If i would get a bunch more tokens to try it i would not complain. Feel its better in writing business models then coding.

Did anyone try to use Gemini 3.1 Pro with Codex as the harness? Is Antigravity the problem with using Gemini for agentic coding, or is it the LM?

@TheRohanVarma when 3p inference in codex ?

Many times a day now, people much smarter than me tell me they are Codex-pilled and that GPT-5.5 was a watershed moment for them. Engineers keep telling me the Codex App is the first interface that got them to leave the terminal agents behind. The Codex App is a fundamental shift in the way I work. I can't even imagine what I was doing before using the Codex App, but it definitely wasn't pretty. Check it out and let us know how to make it even better :)

@basedjensen I hate it to, but is the alternative to ship a bunch of cli-s to normies ?

I have always hated this. I really do not understand people who pushed mcp

@davideciffa huggingface.co/lmsys/SGLang-E… would be interesting if you produce something with 3-4x on t hat hardware.

@YYYYOOOO77 1.5x is a pretty bad speculative decoder, our luce dflash + ddtree has a x3-4 speed up and we are working to make it even better

Fast inference needs heterogeneous hardware, a fast small machine with high tflops, high memory bandwidth and a slower bigger one where hosting the main target model. It is way better than a dgx and the cost is similar

@davideciffa Okay, I understand it now. But you need a great drafter and even then you can expect 1.5 x maybe., what is the base tps on a strix ? So the outcome doesn't look promising. You have to do a lot of work.

@YYYYOOOO77 I mean something else, the main model lives in the slow memory (270gbps), you dont split it. The fast drafters (for speculative decode/speculative prefill) live in the rtx . You can think about it like the poor gpu man version of chips with gigantic SRAM like groq

@Yampeleg It's still a huge mess. too much CPU if you have multiple ongoing sessions randomly using the built-in browser use or trying to get to the Chrome MCP directly. The positive is it creates a follow-ups automatically and the auto-compacting can be great, but also can be a huge miss

I was not expecting the Codex App to be even better than using the terminal. Highly recommend everyone to try. If you are on Linux just tell GPT-5.5-xhigh to “find a way to get it, it’s known to be easy”

International tourism (number of annual arrivals): 🇫🇷 France: 117.1m 🇵🇱 Poland: 88.5m 🇲🇽 Mexico: 51m 🇺🇸 USA: 45m 🇹🇭 Thailand: 39.9m 🇮🇹 Italy: 38.4m 🇨🇿 Czechia: 37.2m 🇪🇸 Spain: 36.4m 🇨🇦 Canada: 32.4m 🇭🇺 Hungary: 31.6m 🇨🇳 China: 30.4m 🇭🇷 Croatia: 21.6m 🇮🇳 India: 17.9m 🇹🇷 Turkey: 15.9m 🇩🇰 Denmark: 15.6m 🇩🇪 Germany: 12.4m 🇬🇧 UK: 11.1m 🇦🇷 Argentina: 7.4m 🇷🇺 Russia: 6.3m 🇧🇷 Brazil: 6.3m 🇳🇬 Nigeria: 5.2m 🇯🇵 Japan: 4.1m 🇮🇩 Indonesia: 4m 🇸🇪 Sweden: 1.9m 🇦🇺 Australia: 1.8m 🇳🇴 Norway: 1.4m 🇨🇺 Cuba: 1m 🇵🇰 Pakistan: 0.9m 🇫🇮 Finland: 0.9m 🇲🇻 Maldives: 0.55m 🇮🇸 Iceland: 0.5m 🇻🇪 Venezuela: 0.4m 🇲🇩 Moldova: 0.03m Source: Yearbook of Tourism Statistics, Compendium of Tourism Statistics and data files, UN Tourism. Data from 2020 or latest available. France is 2020, Poland is 2019 for example.

@skeptrune somebody in his 20s is writing his posts.

alright, it's getting cringe

@AnushElangovan @tbpn Its logitech yeti orb. I have it mapped to, record, enter, ctrl +u(delete), ghostly, chatgpt and alfred. i dont use that program, i use karabiner-elements to remap keys.

@YYYYOOOO77 @tbpn I like that mic. What is that ? The keyboard programer is a virus. I had to rewrite it. What do you map your six keys to ?

AMD's @AnushElangovan modified a Strix Halo to run a local voice model that does real-time voice transcription. It has a 3-key keyboard (next session, previous session, and push-to-talk) and a knob for the model selector. He showed it off on TBPN today:

@bygregorr @minhsmind Nobody is nerfing models, they're just optimizing inference. Nerfing is a side effect.

@minhsmind I'm not sure "compute constraint" is the right frame here. Demand spikes don't nerf models, deployment decisions do. Those are two very different conversations.

Okay so. Let me get this straight. 1. Anthropic’s Dario Amodei battled it out with Secretary of War Pete Hegseth. 2. The result is demand for Claude goes so high that Anthropic gets into massive compute constraint nerfing their best new model: Opus 4.7. 3. The result is users downgrade to 4.6. Not only that, Anthropic cuts multiple services like Cowork as well as changes the terms of $20 per month Pro plan to restrict how much tokens users can use. Essentially, the subsidization era is hitting its first discount constraint. 4. MEANWHILE. OpenAI’s demand goes down because a massive amount of enterprise and consumer users start switching over to Anthropic, leading to a slowdown in the growth that OpenAI was projecting. ChatGPT image 1 Ghibli + Sora2 caused massive spikes in usage in 2025, but that momentum has slowed in the wake of Claude demand in 2026. Sora2 is dead. Image2 is released to some fanfare. GPT5.5 is out but momentum has allegedly slowed according to OpenAI CFO. 5. This puts into question the two upcoming IPOs of both Anthropic and OpenAI. Respectively, can one continue to deliver the results of the last 6 months (making deals with both Amazon and Google for compute to do so) and can the other actually be the Kleenex of AI and the consumer epicenter of AI? Are they both trillion dollar companies? Conclusion: This one event, Dario’s blog, shifted the balance of power in AI. P.S. Google IO is in one month.