Milan

I can genuinely say I'm building what I believe will be called LLM 2.0, although that definition hasn't landed yet. What's interesting is that I already see a roadmap beyond it. The next few years won't be about bigger models.

@XRarchitect @Scobleizer Do is see blue screen bleeding on the side ? 🧐

Testing sol 5.6 tonight. Built this using 8thWall webAR + sparkjs (open source) with a world labs gaussian splat generation of new interior design of my living room

@EHuanglu You will pay royalties to Mr Bean or?

sam altman was caught copying apple in 4k

@dante321321321 @alexeixbt FAFO

@alexeixbt If he tugs on the steering wheel a little too hard in either direction, the car will wipe out. This guy is stupid

when she talks about having kids but my story ends at 23 in my ferrari in the coast of Italy at 330km/h

@MiaAI_lab They paying his company a couple of billions a month. No wonder he changed his opinion. 🙃

How can you not be a fan?

@haydonryan At least someone uses proper server hardware!

It’s alive! The experimental Titan. 192GB VRAM, 192GB DDR4 and PCIe 3.0 to slow everything down lol.

@1337hero How about MI100 1229 GB/s

@The_Only_Signal Dude completely leaves off the AMD RX 7900 XTX and the R9700 just to make NVIDIA look like it has no competition. AMD AI Pro R9700 GPU, 640 GB/s AMD RX 7900 XTX GPU, 960 GB/s Earned an Unfollow with this one.

Someone out there likely needs to see this: M4 Mac Mini, 120 GB/s AMD Strix Halo, 256 GB/s Nvidia DGX Spark, 273 GB/s M5 32 Core MacBook Pro, 460 GB/s Intel Arc Pro B70 GPU, 608 GB/s M5 40 Core MacBook Pro, 614 GB/s Nvidia RTX 3090 GPU, 936 GB/s Nvidia RTX 5090 GPU, 1,792 GB/s

@PutinIsAVirus @0xSero @theo @theo makes a very valid point in his latest video. What he omits is that today’s API access is heavily subsidized. The only sustainable path is pushing frontier AI onto affordable consumer devices-not selling $4k–$40k AI boxes.

@0xSero @theo this is how MOE models work on "consumer" hardware.

GLM-5.2 Nvidia NVFP4 No pruning, no quantising, exact same model card. 110 tok/s single stream. My friend is a genius, holy fudge. Here's the repo, he got full DS4-Flash on a 5090 + DDR5 github.com/kacper-daftcod… at 38.5 tok/s This is a breakthrough.

@0xSero "no quantising" The codebase shows clearly static 2bit compression

It's kind of crazy to think that less than four months ago I bought two old RTX 3090 Turbos to work on patents and our ERP framework for @teqomaterials. While much of the AI community is benchmarking models, building the next to-do app or another Minecraft demo, we've been quietly scaling. Today we're at 24 RTX 3090 Turbos, 8 MI100s, and already looking for more compute for our PoC. We're also burning a crazy amount of API tokens every day. Good things are ahead. 🚀🇪🇺🇵🇱

@eli_lifland This assumes every nation will voluntarily give up its strategic advantage. That's not a plan. That's wishful thinking. AI sovereignty isn't built on hope-it's built on technology. Fortunately, we're already building an alternative in our lab.

Excited to finally publish AI 2040: Plan A, our plan for international coordination to get to a great AI future. We've put a lot of work into this and I hope that it sparks alternative plans that bring us closer to treating superintelligence with the seriousness it deserves.

@ClaudeDevs Sweet

We've reset 5-hour and weekly rate limits for all users.

Seems I'm not the only one who'm returned to Twitter after a decade due to AI... 🤔🙃

@deZeZed Why not a tuned Audi or Porsche? Mustangs are only for straight roads !

Zdjęcia: echodnia eu

Szybko poszło, już skasowali Mustanga Policyjnego xddddd

@DavidOndrej1 No worries David, we'r working on an EU alternative teqolabs.eu

YOU MUST BE JOKING

@norbertbodziony teqolabs.eu soon you'd be running from our stack 🙃

Who in Poland is actually building with AI? Not just talking about it. Building products, agents, automations, workflows, internal tools. If that’s you, comment here or DM me. Let’s talk.

@_ptonik @0xSero Finally someone who can do that properly. Can't look at those hacked MacGyver setups anymore!

@0xSero i'll venmo you $1k right now to let me in and fix this cable management. my sleep paralysis demon routes cables like this attached is cable management that goes to heaven. yours is getting left behind in the rapture

Friends in high places.

@antirez @1337hero @theo What he goes beyond to, is that API costs are overly subsidised. If you'd compare 6x ~ 8x Tenstorrent p150 with the real API costs, you'd be brack even within the month. This subsidising wont be taking that much longer in my opinion, Fable pulling out the max sub is a good e.g.

The point that @theo makes is that it is rare to run DeepSeek v4 Flash on local hardware, you need special hardware that people don't have, and even like that, it is slow-ish, even if usable depending on the setup. For instance: what is your speed at largish contexts with your setup? It will be good enough for many use cases, but working on large codebases fast will require 150 t/s generaton, 4000/5000 t/s prefill. With this numbers it could start to look like AI provider experience alike. Even like that, it is not going to be GPT 5.6 or Fable, of course, but can approach Opus-class models (without vision). You see how many issues? But yet: it is the path forward.

Theo: local models = "a quantized Gemma 4 that barely functions." I'm running Gemma, Qwen Coder, and DeepSeek Flash on 3x AMD R9700s (96GB, $3,998) My codebases disagree.

@Scobleizer Is this what an AI psychosis looks like ? 🙃

Thank goodness for AI alignment.

@ajaeger74 @LLMJunky Cudacores too.

@LLMJunky What matters is bandwidth

Recent leaks show the upcoming RTX 5080 SUPER is rocking with 24gb of 32GB/s DDR7 VRAM. A bit faster RAM than 5090, but 256-bit bus so still slower overall. But still, at an estimated ~$1300 price point, this will be a better alternative to the now dated 3090 >higher bandwidth >blackwell >didn't mine bitcoin for 6 years before you got it) Not a bad compromise if you're on a budget at first glance

@LLMJunky Without NVLINK, the 3090 turbo is still the better option.