Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)

@teortaxesTex on X

Writes about AI and Engineering. 82% of posts open with an observation. A typical post reaches 339 likes.

Remix their style

Posts: 33
Likes: 16K
Typical: 339
Best: 2.5K

What they write about

AI: 76%
Engineering: 21%

How they open

observation: 82%
number: 9%
question: 9%

Every post of theirs in the library

Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex

I really admire them. I respect the projects of other open teams too, but NOBODY comes close to DeepSeek's purism. For them, this really is the whole vision. No subscription promos, no clauses, no announcement at all. No mind games. «Intelligence. Too cheap to meter. Today. MIT»

Opens with an observation

Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex

> $5.5M for Sonnet tier it's unsurprising that they're proud of it, but it sure feels like they're rubbing it in. «$100M runs, huh? 30.84M H100-hours on 405B, yeah? Half-witted Western hacks, your silicon is wasted on you, your thoughts wouldn't reduce loss of your own models»

Opens with a number

Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex

I have tried to use OpenRouter Fusion API with cheap open models only, and saw reasoning that surpasses any of them individually. Then I looked into API logs and saw that this "Fusion" still calls Opus 4.8 as a judge. I see no way to disable it. Not cool, OpenRouter. Not cool.

Opens with an observation

Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex

DeepSeek: «We plan to raise the overall pricing for DeepSeek API services in the near future, with a significant increase expected. Please plan your usage accordingly» Oof. They originally promised price *cuts* in H2. given the market situation… they are heavily overloaded.

Opens with an observation

Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex

listening now Xi makes a speech that's basically DeepSeek recruiting copy. AI as steam and electricity. «We human beings must answer the question posed by our times. How to get along with thinking machines?». Open source AI for all humanity is part of Xi's answer. Buckle up.

Opens with an observation

Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex

> All benchmark results are reported under INT4 precision. Do you understand what a flex this was. They go toe to toe with GPT-5 on the heaviest, longest-range tasks, with hundreds of tool calls. ALL IN INT4. «Convert to fp8 if you need» Frontier lab.

Opens with an observation

Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex

Junyang had been announcing Qwen models for… not so long, 2 years? Yet it was more models – many of them absurdly strong – than a dozen other labs together have released in their lifetime. I am a DeepSeek fan, but I have to admit: Qwen is what largely kept open source alive.

Opens with a question

Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex

VP of Science at OpenAI hyped up that GPT-5 «found solutions» to 10 open «Erdös» (actually Erdős but whatever) problems, got fact-checked by the literal owner of erdosproblems dot com, deleted the post. AI might not be a bubble, but OpenAI bros live in an echo chamber

Opens with an observation

Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex

after giving Manus a spin I conclude it's a product devilishly optimized for influencers, which is why it exploded so much. Generating threadboy content, trip plans and such general interest 🤯👇 stuff – yah. STEM assistance, coding – worse than googling. More LLM than agent.

Opens with an observation

Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex

DeepSeek-V4-Flash is updated. «DSBench-FullStack is an internal full-stack development test set, and DSBench-Hard is an internal Coding Agent hard-problem test set» «The official release of DeepSeek-V4-Pro will follow soon» These scores are… quite a lot better than Preview.

Opens with an observation

Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex

Massively unexpected update from DeepSeek: a powerful, high-compression MoE OCR model. > In production, DeepSeek-OCR can generate 33 million pages of data per day for LLMs/VLMs using 20 nodes (x8 A100-40G). They want ALL the tokens. You're welcome to have some too.

Opens with an observation

Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex

GLM 5.2 is apparently the strongest Chinese model, and it'll be open this is quite a jump on CritPt. They have *finally* started to build general reasoners. I think it'll be very competitive on ARC-AGI 2 and WeirdML, too

Opens with an observation

Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex

anon do you realize that V4-Pro is straight up the strongest pretrained model we have? Like… 1.6T@49AB (≈280B dense), 33T – even by meme formula it's > LLaMA 3. Add Muon, mHC, most steps 64K context + extended to 1M… No excuses now. Every "unicorn" can have its brand AGI.

Opens with a question

Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex

…It wasn't an intern's joke MiMo 2.5 (not Pro): > Trained on a total of ~48T tokens using FP8 mixed precision. The context window supports up to 1M tokens. We've got another 1M class, and the largest disclosed pretrain. Congrats Xiaomi.

Opens with an observation

Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex

Western open source frontier is back. This is on par with the largest of known Chinese pretrains so far (though might be outclassed in a few hours). And yes, it's another DSMoE +trained on Kimi data. But it has vision and *audio*. Unexpected good news.

Opens with an observation

Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex

Do you get what this means anon [nerfed] GLM 5.2 can materially help in defending against an absolute private frontier model above 5.6 Sol that's not nerfed on cyber and is autonomously attacking. How do you think Dario's plan to pwn the CCP will go

Opens with an observation

Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex

on the contrary, Huggingface should freak out about a world where OpenAI and Anthropic can fuck you up and agree to deny you any means of defense. Natural slaves will find this situation acceptable, perhaps, but it's really creepy

Opens with an observation

Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex

This is important too Two more notes: - they don’t call models “exp” or “preview” as a joke. 3.2-exp was half a generation behind 3.2. Vision-Full will go crazy - nothing said on open weights so far

Opens with an observation

Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex

> built through continual pretraining on approximately 15 trillion mixed visual and text tokens atop Kimi-K2-Base …It's essentially a totally new model with new abilities. 30T tokens @ Muon. «Kimi K2.5 represents a meaningful step toward AGI for the open-source community» wow ok

Opens with a number

Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex

so what's the justification for openly releasing Opus 5 if it's de facto the strongest Anthropic model we know about (Fable 5.1/Mythos-New are rumors about internal progress and don't count), and the strongest model period?

Opens with an observation

Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex

the power of scaled RL + distillation (but big boy distillation, not cringe SFT on claude subscriptions) Anthropic is ahead, everything else is cope. OpenAI can win yet but 5.6 is no evidence of anything except post-training expertise. They'll need bigger models

Opens with an observation

Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex

API is overpriced vs subscriptions by 40-70x Do you know what this means? They have DeepSeek-level compute efficiency if not better. Which makes perfect sense (tbh still too good for DS), given frontier budgets for overtraining, autoresearch, larger and older staff etc.

Opens with a number

Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex

Grok seems to get tired and start lazily cheating later into the session (247K/500K), or I just got unlucky in any case, the sheer speed and decisiveness of the model are amazing, this is what Gemini 3.5 Flash was supposed to be

Opens with an observation

Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex

This is also what @ylecun means when saying that GPT-4 or "GPT-5000" doesn't have a cat's understanding of the world and can't anticipate simple physical causal chains not present in texts. It's easy to mock his dismissive takes – but, incredibly, there is evidence behind them.

Opens with an observation

Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex

Remember: @deanwball doesn't write USG policy anymore, but he has a fairly good idea of how it can develop. It doesn't much matter if he approves. Dunking on Dean in his OpenAI employee capacity is fun. Is this fun too? Well, *I* am having a lot of fun.

Opens with an observation

Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex

Oh well. America Winning once again. Wypipo innovate, the Chinese copy. I feared that it'll take exactly this clownish, reactive form of "they did it so uh we must too". Confirms that the CCP doesn't have any serious independent thinking about AGI.

Opens with an observation

Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex

Ok these are pretty good scores …though some crucial agentic ones (NL2Repo, DeepSWE, Agent's Last Exam, AutomationBench) are surprisingly close to the new V4-Flash. But it has vision and it's a big boy. Also, for the first time ever, we'll see @htihle test Qwen-Max. interesting.

Opens with an observation

Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex

Defeated with cheap rewriting, definitely defeated if you train a rewriter against a reconstructed detector. Amusing GAN project actually But at least will help Pangram to deal with the low effort slop

Opens with an observation

Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex

So speaking of benchmarks, what can be said of the new open Qwen? First, it completely destroys Qwen3-VL-235B ofc, but more surprisingly it outscores Qwen3-Max-thinking. All the while it's the same model as "Plus". Plus just has 1M context and some more bells and whistles.

Opens with a question

Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex

> The depth of the computation graph for our present frontier models, including Astra, is within a factor of two of GPT-4. so <240 layers? Makes sense, hardware punishes anything deeper, and training becomes harder too… But… been a while since we really Stacked More Layers.

Opens with an observation

Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex

I'd actually say that "open source" as a dumping strategy makes little sense The point of dumping is to bankrupt the competition, yes, but you're still selling Chinese OS, in theory, just shifts the revenue from US labs to US neoclouds. That's… not a big blow to the US

Opens with an observation

Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex

> We propose IndexShare, which reuses the same indexer across every four sparse attention layers, reducing per-token FLOPs by 2.9× at a 1M context length. We also improve GLM-5.2’s MTP Good demonstration that a lot could be squeezed out of "mere" DSA on benchmarks, too.

Opens with an observation

Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex

DeepSeek is trolling They don't say *what* they trained on, only "eh MXFP4" Hawks will now go on a hunt for Blackwells in Inner Mongolia In a year they'll publish "insights into V4" with details of an experimental Huawei cluster man I love this lab so much

Opens with an observation