Sourav

Deepseek shipping speed is truly underrated. In just 16 months, they shipped 10+ architectural innovations across training stack. > MLA > DeepseekMoE > Aux-loss-free load balancing > GRPO > R1-Zero > DSA > CSA (in v4) > HCA (in v4) > mHC > Large-scale agentic task synthesis and almost everything is open source. what am i missing?

13!

Why is no one talking about this? @nvidia is offering around 80 AI models via hosted APIs absolutely for free. You get access to MiniMax M2.7, GLM 5.1, Kimi 2.5, DeepSeek 3.2, GPT-OSS-120B, Sarvam-M etc. This plugs straight into OpenClaude, OpenCode, Zed IDE, Hermes agent and even with Cursor IDE. Setup: – Grab API key: build.nvidia.com/models – base_url = "integrate.api.nvidia.com/v1" – api_key = "$NVIDIA_API_KEY" – select model (e.g. minimaxai/minimax-m2.7) If you’re building or experimenting, this is basically free inference. Lock in and start building today anon. Thank me later.

Ramsey numbers are one the most basic objects in combinatorics, a beautiful illustration of structure within chaos. They have been heavily studied for almost a century now, so it came as a real surprise to us when an internal version of GPT-5.5 proved a new elementary result about them: \lim_{n\to \infty} R(k,n+1)/R(k,n) = 1 for all k This was also known as Erdos problem #1014, although I personally think the more relevant bit is that it's a basic result about off-diagonal Ramsey numbers. As it often happens (for now), the proof is reasonably simple in hindsight, although it is quite a wire act and it relies on some unexpected numerics (the "unexpected" part here is probably why this wasn't discovered before). Despite being simple, it's certainly the type of result that could now be taught in a combinatorics class. Pdf of the proof (produced by @mehtaab_sawhney): cdn.openai.com/pdf/6dc7175d-d… Lean verification of the proof (produced by Boris Alexeev): github.com/plby/lean-proo…

DeepSeek V4 hits it out of the park and addresses HBM shortage: DeepSeek proves why it is such a fundamental research lab. In addition to exceeding Opus 4.6 on Terminal Bench and virtually matching on other performance metrics, the most notable advancement is this statement: "In the 1M-token context setting, DeepSeek-V4-Pro requires only 27% of single-token inference FLOPs and 10% of KV cache compared with DeepSeek-V3.2" To understand significance of this point, consider below diagram that shows memory layout for Prefill and Decode nodes. If you implement Decode with Data and Expert parallelism (DEP16) with 16 GPUs on GB200 or GB300 NVL72 rack with DeepSeek v3.2, you are left with 104GB or 176 GB HBRAM per GPU respectively. Here we are assuming MoE parameters are in NVFP4. The remaining HBRAM per GPU dictates how large batch size you can have for inference, which determines how many concurrent request you can serve. Consider GB300 with 176GB left: 1. For 128K context, you need 4.45 GB HBRam for KV Cache, and you can serve only 36 concurrent requests. 2. For 256K context, you need 8.90 GB HBRam for KV Cache, and you can serve only 18 concurrent requests. 3. For 512K context, you need 17.80 GB HBRam for KV Cache, and you can serve only 9 concurrent requests. 4. For 1M context, you need 35.60 GB HBRam for KV Cache, and you can serve only 4 concurrent requests. You see the point. Now you imagine, you actually required 10 times less KV cache somehow at 1M! It basically enables you to server 10 times more requests with same resources. Recall Decode is memory bound and not compute bound, unlike Prefill. This is probably the most important contribution of DeepSeek V4. @teortaxesTex @jukan05 @zephyr_z9

The most underrated LLM afaict. Kimi works really really well too on @windsurf Looking forward to trying it there

The 2.6 benchmarks are closing in on Opus 4.6. Dayumn

@ml_angelopoulos The evidence seems to align as well x.com/i/status/20460…

@ml_angelopoulos The pricing shock is definitely upsetting. I can get a lot done by 4.6 without running out of my session limits. But 4.7 is a real guzzler

To be clear, 4 points over Opus 4.6 is nothing to write home about, and is not significant according to the CIs. The vibes on X have been correct about this model so far, in that it is different from, but not uniformly better than, Opus 4.6. Still, Opus 4.X remains king 👑

@PMinervini @iamtrask You are in an enviable position of having both :)

@srvmshr @iamtrask Got an m3 max as a laptop and sometimes miss not being able to run cuda software right away, the spark fixed that very nicely

Hey - if you don’t use a local LLM as your daily driver, why? What are the remaining problems?

How does one delete tasks in @OpenAI Codex on Web? I see only options for Archiving. Can anyone chime in why this feature is missing, especially even on Teams plan where we opt out of data retention @OpenAIDevs @thsottiaux

@iamtrask Spark is more tinker-friendly & a big W if writing anything remotely low level (CUDA, fine-tuning, kernels etc). Otherwise yes, MLX + GGUFs makes it pretty convenient

@srvmshr Seems like the Mac Studio van at least be a great Mac if it doesn’t work out, yeah?

Gut punched @projecthailmary tumblr.com/wolfythewitch/…

@thsottiaux @barret_zoph Please introduce "Delete Tasks" on the codex web UI. Frankly, it keeps code information even from disconnected GitHub accounts/repo in the task panels. Our org doesn't share data for model improvement. This isn't okay. People report that some github issues pending around it.

Codex Compute efficient ✅ Always up, never down ✅ Best at hardcore engineering ✅ Crazy good app, first to escape the terminal ✅

@tarantulae Model distillation is quite likely. Opus 4.7 is downstream of Mythos. I have some serious concerns about a gimped version of the larger, more capable model i.e.more token usage but variable quality

Interesting that Opus 4.7 seems to be using a new tokenizer, which means probably a new model, but Mythos is the new model, so I wonder what is going on 🤔

@liztansz Find the sessions you like & walk around to posters that interest you. The Q&A usually tells things that aren't in the paper. The socials in the evening are nice. Outlier tip: nice to have business cards if you talk to people. Food usually is mehh - don't count on it

can people please share their top conference tips ? Im going to ICLR next week !! do people prep for conferences?? what do you do there? do you just go to make friends ? please-- this is my first conference !!

Very fair points. François makes the correct distinction that scalability isn't a property of simplicity. I think scalability has dependency on how well the solution space can be parallelized & also parametrized. Probably what coincides with his views on hi-entropy

ML folks, esp. the mass publishing usual suspects, rather believe otherwise