Kunal Tangri

Claude’s constant outages are actually intentional. Part of safe AGI roll out is making sure we’re not 100% dependent on it. Only 98.94% dependent

Anthropic rejecting the DoW has generated a lot of good data for altruistically aligning their next model

💥💥💥 New(-ish) paper alert 💥💥💥 Metient is a fast and powerful method to reconstruct the migration history of metastatic clones and learn cancer-type specific patterns of metastatic spread, from bulk or single-cell sequencing data. nature.com/articles/s4159…

I've spent the last few years interviewing 100+ AI engineers at Farsight and advising Fortune 500s on AI strategy at Hugging Face. One pattern keeps showing up, and it's gotten worse in the LLM era: Overcomplication and underevaluation Great systems start simple. Then, evaluation is what drives complexity. Today, too many teams chase shiny new frameworks, tools, and model capabilities instead of focusing on what matters - is the system actually working better.

Maybe not in its current form, but never is a strong word

GPT-5 in ChatGPT is meh GPT-5 in Cursor is incredible. Just did something way more complex than I could've with Gemini or Claude I tend to be very detailed in my prompting but very disorganized. GPT-5 feels like it just gets it (in Cursor)

@NielsRogge Yeah router definitely feels broken. Need to put think really hard in most prompts now

Interesting, so this is a router issue. It should learn that queries like this require thinking Would the router be a BERT-like model? An LLM?

@Eriks_Life @sama Not the quality, it’s literally broken when I try to use it

@TangriKunal @sama Wouldn't it be fixed automatically with Gpt-5? Since it was the version on Gpt-4 that wasn't that good.

GPT-5 now rolled out to 20% of paid users and doing >2B TPM on the API! so far so good... excellent work by the eng and infra teams!

Did the GPT5 launch break agent mode for anyone else?

@jxnlco Not quite UES but we’re on 53rd and 3rd

anyone have offices near the UES?

So proud to share that Farsight has announced our Series A! Over the past year, we've seen incredible traction with leading investment banks, PE firms, and hedge funds who are using our platform to automate everything from deal sourcing to complex financial modeling. We're at an exciting inflection point: scaling rapidly to stay ahead of the fast-moving AI landscape, constantly fine-tuning the delicate balance between accuracy and adaptability. This positions us uniquely to rapidly integrate emerging AI technologies into practical, impactful solutions for our customers. Grateful to our investors who believe in our vision, our customers who trust us with their most critical workflows, and our talented team who makes this possible. 𝐖𝐞𝐛𝐬𝐢𝐭𝐞: lnkd.in/evKASUu6 𝐏𝐫𝐞𝐬𝐬 𝐑𝐞𝐥𝐞𝐚𝐬𝐞: lnkd.in/eHiqG8jk

So much competition at the pure intelligence frontier. Not enough going on at the optimal cost/intelligence frontier

@Hans365days @TheAiGrid Tbh found that it’s not very good at coding. Much more strongly suited to research/multistep function calling

@TheAiGrid I will try o3 in windsurf today. First time using o3. Any tips?

OpenAI really messed up by calling the model o3. It should be called Agent 1. Because its literally an agent that can go off and do things for you. Not just a chat model that you talk with. And people arent realising how powerful the model is because they are treating it like a chat model..

Guess MCP just did it way better 🤷🏾‍♂️

@RileyRalmuto Current LLMs rely on compressing human-generated data - upper bounded by human intelligence. I think Ilya saw a path to synthetic data generation guided by utility rather than imitation

remember how I told y'all that Ilya's been working on an entirely new form of data? and a large part of the reason for their silence (aside from not having a product or need for marketing) is due to the nature of this work? not too exciting on its own (to the general pop) ya know? but when it is time...you will see. what do you think Ilya saw, with this information in mind?

@aidan_mclau What if you could also set limits on the recency of info deep research can use? Could see how those probs evolve over time

My Twitter algo feels way better today than it has in months