Thomas Ip

@elonmusk every time you open sourced the algo, you didn't release the params and weights. totally useless. it's like open sourcing a LLM without the weights.

To give people confidence that we are not secretly manipulating the 𝕏 recommendations, it is critical that we open source anything that influences what people are shown

@AstravoreBC @w_milczynska Most of this is not UK specific.

@w_milczynska 🤣🤣🤣🤣🤣

UK founder starter pack (save this): → incorporate via Companies House: £100, 30mins → bank: Mercury (US)/Starling (UK) → legals: SeedLegals (~50% new UK startups) → EMI BEFORE first hire → HMRC advance assurance pre raising → cap table tools from day 1 a weekend, <£1k

@UnsocialB86776 @Hesamation it's the opposite way round. 4o shipped, thinky is a research preview that is not likely going to be launched.

@Hesamation GPT-4o demo was impressive but Thinking Machines shipped something people can actually use. Demos don't win, products do.

WAIT WAIT WAIT. is anyone gonna talk about the fact how Thinking Machines demo looks insanely similar to GPT 4o demo from 2 YEARS AGO?

@VraserX 4o voice is a massive production model with turn based interaction. Whereas thinky's model is "full duplex"/micro turns and a much smaller research model. I think the gap is simply from an unclear demo and a very early preview of their work.

I’m not really impressed by the Thinking Machines voice mode demo. It feels like they basically recreated GPT-4o voice mode, and the comparison is impossible to unsee. It is even the same researcher in the demo. Yes, full duplex is a nice upgrade. But the responses feel slower, the voice feels less natural, and the whole thing has less magic than GPT-4o did two years ago. OpenAI GPT-4o VS Thinking Machines

Superintelligence will be achieved by whoever takes the bitter lesson the most serious: A single all-to-all MLP with no human shortcuts, RL'ed in a equally bitter-lessoned world model.

@LaurenceBrem @miramurati that is not that fast? any of current voice models like chatgpt voice or grok voice have the same latency. thinky is different something different though, looks like they want to handle input, output and tool calls at the same time.

@miramurati How did you get the voice model to respond so quickly? 🤯

Today we're sharing our work on interaction models. A new class of model trained from scratch to handle real-time interaction natively, instead of gluing it onto a turn-based one. youtu.be/A12AVongNN4

@_architected @miramurati interruption is still turn based. i think thinky is trying to train a full duplex model that can take input, produce output and run tool calls at the same time. they didn't demo the full duplex between input and output though, only tool calls and output.

@miramurati don’t they already have interruption mode in Gemini? forgive me if i’m stupid but how’s it different for the end customer? i do under what you’re trying to achieve but if it looks and feels the same, what’s the point?

Longest codex /goal run yet at 22 hours! I am optimizing Qwen3 Omni on SGLang to be able to serve 30+ concurrent real time voice chats on an AMD MI300X. Starting with one active chat not even hitting real time, it's almost reaching the concurrency goal after almost a day of optimization and profiling. I thought it was cool how people were able to get agents to run automonously for a long time, turns out this is actually boring and unproductive. Still a long way to go for LLMs to get so good they can think for a few minutes and oneshot the solution. Make no mistake, the frontier models are very much AGI, but I can't wait for superintelligence to come fast enough.

@aschmelyun skills is basically functions but for llms. not using skills is like writing your software all in the main block

Feel like I'm missing out because I don't use skills, or a lot of MCP, or multi-agent orchestrations when using AI dev tools. I'm just like "implement this feature" or "how do this work" or "no not like that, do this instead". Idk, I feel fast and accurate so why change?

@kwindla No tool calling unfortunately, the instruct/reasoning is just for conversational use case. The audio output will get messed up if you add tool calls output. I don't think there are open speech to speech models that support tool calling or even reasoning? (dont quote me on this)

@_thomasip Have you fine tuned Qwen 3 for long multi-turn use cases with tool calling? Things like what the aiewf-eval benchmark tests. Would love to see examples.

This is a great question! There are not (yet) any speech-to-speech models that are workable bases for production fine tuning. But that's likely to change pretty soon. To be a good starting point for fine tuning, a base model needs to already be reasonably good at multi-turn instruction following and tool calling. My mental model, here, is that the fine tuning process needs to be able to "find", in the model weights, the conversation patterns in your data set. If the model is too weak, those patterns aren't there to find and emphasize. My experience fine tuning text models for multi-turn and voice agents, this means the model needs to score something like 85% on the aiewf-eval benchmark. So today that's text models like Nemotron 3 Nano, Gemma 4 31b, Qwen 3.5 27b, GPT OSS 120b. There aren't any open weights speech-to-speech models that are close to this bar, yet. The best options today are the Kyutai Moshi and NVIDIA Personaplex (which is a Moshi-family model). Moshi is a very nifty model architecture. Fine-tuning these is a great research project. I don't think you'll get to production level performance for most voice use cases, though, no matter how much fine-tuning (or even post-training) you do on these models. There's a new NVIDIA Nemotron Voicechat model coming soon, though, that should be a great base model for voice agent fine-tuning. It's in early access. You can try out the demo ...

Qwen3 is quite highly regarded and popular for fine tuning, what's your problem with it? And yes 3.5 omni is not open, and sadly unlike to happen. Open omni models are rare to come by. The only issue i have with 3 omni is its pure attention arch unlike 3.5 so it's quite kv cache hungry to serve in production.

@_thomasip I didn’t have a great experience with Qwen models until 3.5. I think weights for 3.5 Omni aren’t released yet?

@AssetOfBaskets @paulbohm you are mistaken

@paulbohm This must be a fake history by the way. Nobody is that retarded

If your startup does not have a UUID microservice you’re ngmi

@dejavucoder people will do anything but look at the damn data

everyone wants to do RL, no one wants to look at their prompts and data

@birdabo japanese X is unhinged

never deleting this app 💀

@teortaxesTex wonder how much engineering ideas are exchanged between deepseek and high flyer. access to the best quant engineering in china is a big moat

Superiority Yes, this is DeepSeek cache hit statistics, the horizontal bar at 100% As I've said, they have a strictly optimal implementation. If there is a context to be reused, it will be. I think 12 hours is too short a window, too. It's probably 24-48

@Ed_of_O @paularambles what's the second one? Nvidia?

@paularambles Everything's in a private campus and he's not a member.

just met a japanese guy visiting sf for the first time who was deeply disappointed by how not-high-tech the city was. paris syndrome but for people expecting san francisco to look like the future

@ludwigABAP wait deepseek doesn't instill moral judgement and refuse requests? Might genuinely be useful as a companion to claude/codex

"deepseek-v4, reverse engineer this and get thru the paywall, go" 20mn later

@alex_whedon @bomboraassclaat you outsourced your graph making??

@bomboraassclaat What would you want to see in the paper? I apologize for the graph! We outsourced it, and I didn't catch the disproportionality until someone pointed it out on Twitter. Definitely not intentional!

Hey, folks! We have been blown away by the response to SubQ and the SSA breakthrough over the last 48 hours. It is awesome to see how many people are responding to our mission of creating more efficient algorithms to create better models. We are working hard to firm up our release timeline and will share more very soon. We will also share additional data and third-party validation in our model card next week. If you have questions, please post them in the thread, and I'll do my best to respond! Above all, THANK YOU! The support, feedback, and discussion from this community have been inspiring.

@Joedefendre @flowersslop grok 4.3 upped the nsfw refusals, they are going corporate too.

@flowersslop How do you feel about the grok voices ?

Only hopium left is that the realtime 2 model is corporate clanker like on purpose so companies like it better but then in chatgpt they gonna make it really personal and human and warm and natural we will see