Strait of chingrighata finally open.
Sreyashi Dey@SreyashiDey
First in my bloodline to see #ChingrihataMetro construction work 😍
English
Rahul
1.4K posts
Rahul
@guddur
Video games, East Bengal, Chelsea, Machine Learning @ a fmcg
Bengaluru Katılım Haziran 2009
198 Takip Edilen59 Takipçiler
Rahul retweetledi
Rahul retweetledi
🚨 BREAKING: Italy are OUT of the 2026 World Cup.
Third World Cup missed in a row. ❌🇮🇹
English
Launching something we’ve been working for a while - AI agents to build chips faster.
Applied to YC. Let’s see where this goes 🚀
archgen.tech
hari_haran@HariAyapps
@naveen_venk and I submitted our @ycombinator to build the next generation of AI chips with agents. Do you think we'll hear back? @garrytan @snowmaker @t_blom
English
@shantanugoel I will prefer reranking + semantics any day but actual data we deal with is shit. And don't have option to train our own embedding, which would solve 90% of issue. Long run much less token costs and better retrieval. But heh!
English
@shantanugoel Yes. But only if your doc are clean. I tried this a memory for a agent I am working on. Failed horribly because the it is getting bs from vector search as it contains lingo that make no sense to the model (bad score).so moved back to this. I guess it work great if self train emb
English
Also trying an experiment with oxydra which renders "SKILLS" obsolete as mere context bloat.
I'm trying a mode so when the model tries something new and makes errors and finally stumbles upon the right way, it would remember in long term memory what fails and what works, thus upskilling it self.
With the awesome vector search baked into libsql, it can easily recall its past skills without bloating big skill sets into its context unnecessarily.
Shantanu Goel@shantanugoel
Updated oxydra last night so that it adapts to model capabilities. - If the model has ability to parse audio, it will send it to model directly otherwise the model will try to convert it to text (not baking in special tools by myself) - same for docs/images/videos etc
English
@shantanugoel Another thing I found out building and storing a function by agent itself and write to file. And later recall that function from a function registery. Saves a lot of token on regular tasks. Where you don't have stored procedure/function.
English
@shantanugoel You ideally can have subagent that gets triggered on failure with a job of searching skill and feed tha back to main. Generally very little context bloat here. But again depending on how you are extracting and storing skills. My vector search failed horribly because shit data stc
English
Man it took a lot of effort write those books, codes for training 😌
Anthropic@AnthropicAI
We’ve identified industrial-scale distillation attacks on our models by DeepSeek, Moonshot AI, and MiniMax. These labs created over 24,000 fraudulent accounts and generated over 16 million exchanges with Claude, extracting its capabilities to train and improve their own models.
English
@shantanugoel As a side effect of llm being markdown maniacs the documentation became standardized. Took a few years 😌
English
Remember when having these md files in your github repo was a taboo, and people used to gitignore *.md 😁
English
Rahul retweetledi
You can now train MoE models 12× faster with 35% less VRAM via our new Triton kernels (no accuracy loss).
Train gpt-oss locally on 12.8GB VRAM.
In collab with @HuggingFace, Unsloth trains DeepSeek, Qwen3, GLM faster.
Repo: github.com/unslothai/unsl…
Blog: unsloth.ai/docs/new/faste…
English
Rahul retweetledi
Me defending my O(n^3) solution to the coding interviewer.
English
@IndiGo6E wasn't the flight fare was capped by going. Or you got an extention that @DGCAIndia
English
Frozen solid out there 🥶
Squad up in the cold-weather fits. Winter Offensive arrives tomorrow in #Battlefield6 and #REDSEC.
Don’t miss the comments 👀 there’s something buried in the snow ❄️
English
Rahul retweetledi
We disrupted a highly sophisticated AI-led espionage campaign.
The attack targeted large tech companies, financial institutions, chemical manufacturing companies, and government agencies. We assess with high confidence that the threat actor was a Chinese state-sponsored group.
English
Rahul retweetledi
Training LLMs end to end is hard. Very excited to share our new blog (book?) that cover the full pipeline: pre-training, post-training and infra. 200+ pages of what worked, what didn’t, and how to make it run reliably
huggingface.co/spaces/Hugging…
English
@ICICIBank_Care how can suddenly the account type and status can be changed, without informing the customer. How do you even manage this
English