Ankith 🐋/acc
@dhtikna
Herding distributed systems by day and LLM obsessed by night. If you knew what our HFT was cooking youd be jelly 😏
Check my Highlights tab for LLM-only stuff!
402 Following 2.1K Followers
What kind of chart crime is this? If OpenAI or Anthropic did this, you’d all have 5 posts (rightfully) calling it out. But since it’s about him, you don’t even read it, just suck his dick and post "he won." Holy cringe ass Elon fandom
this is how its supposed to look like btw:
Show more
Far more comforting to believe than architecture issue but they also could be two sides of the same coin
Um is lack of compute the explanation for every shortcomings of V4 pro??
> this gain per run
IT'S STILL UNDER-POST-TRAINED
Do you get it anon?
We ran DeepSeek v4 Pro 0813 on our cybersecurity benchmark, it outperformed EVERY (!) other model at finding vulnerabilities
- At pass@3, it rediscovered 87.5% of the benchmark CVEs. Far above Opus 5 and Qwen 3.8 at 81.3%
- The tradeoff is precision. Only 65.6% of vulnerabilities reported by DeepSeek were valid. Far below GPT-5.6-Sol's 86.4%
- The model can also be unpredictable. It only finds an average of 58.3% of vulnerabilities per run. It is strongest when combining its runs' findings.
@deepseek_ai is amazing. They just outperformed every other labs with an open model
1/3 🧵
Show more
No I do think wenfeng is right and continual learning is the next unlock.
General intelligence is not sufficient to excel in a niche domain. Unless you learn the domain (in weights) you will always end up doing (and repeating) search/exploration which is (exponentially) time consuming and non deterministic
Md files are caching on top of search and exploration but again end up consuming valuable context lenth and move search from environment onto context length
Show more
No I do think wenfeng is right and continual learning is the next unlock.
General intelligence is not sufficient to excel in a niche domain. Unless you learn the domain (in weights) you will always end up doing (and repeating) search/exploration which is (exponentially) time consuming and non deterministic
Md files are caching on top of search and exploration but again end up consuming valuable context lenth and move search from environment onto context length
Show more
deepseek v4 pro (0813?)
self-potrait
via midjourney v8.2
Deeply insulted by my X timeline showing me Grok 4.6 ahead of V4 pro GA
"charlie kirk" was probably a mistranslation causing a few historical figures to be merged into one mythological individual. texts from this era are sparse but "charlie kirk" likely never actually existed
Show more
Dammit I'm going to run out of my Fable weekly quota for the first time this week
Can 220° FOV be achieved in a small form factor? Yes!
Together with @HyperVisionXR we are now able to shrink whole electronics
☑️ 4 x 4k Micro OLED displays
☑️ 220° seamless crease-free FOV
☑️ Can easily run on 5070 GPUs or higher
☑️ Wired PCVR
❓ Quite expensive though 🤷
Show more
what the hell
Is there any evidence big model smell / common sense / intuition / grokking of big models can be distilled?
fable 5 is the only usable anthropic model
they should just get rid of all the other ones
Don’t believe anyone saying “Doug” is OpenAI’s new model pretrained from scratch and that it’s coming after Astra.
What actually happened is that SemiAnalysis got that information about the Doug model in early July, before the more recent reports about Astra, and ended up mixing things together in an August 6 article.
Doug showed up BEFORE Astra in the reporting, not after. So there’s no way to confidently claim anything, because Doug could even be Astra under an earlier internal name.
If you’re confused, just know that OpenAI used the name Doug, then Mewthree and MewFour, and only later Astra. It’s still completely unclear which one is which. When I have more information, I’ll share it.
Show more
Oh no they did it again!
luna is such a special model, incredible price performance
@PolymarketMoney no. quantum computing is where quantum computing was 5 years ago
Super jail
No zoomers understand their own work anymore. The prevailing attitude in my technical circles among young engineers is that if you're not straining your understanding on your work you're not getting enough leverage from the models.
I think they're right btw
Show more
Very nice but Grok 4.5 was left out for a reason!
54 @ $0.35
