Mick Gupta

Product @ Brex, SoFi, Meta, AirBnB, Msft, Amazon & Bain - I believe that we need to adapt our philosophy to our science, and not our science to our philosophy

[Brian Armstrong @brian_armstrong]

Finally, the web has native monetization. Ads were not the only way.

[Mick Gupta @Mickgupta]

A new, improved, unsaturated coding benchmark!

[Mohit Raghavendra (@ ACL) @mohit_r9a]

SWE-Interact reimagines SWE benchmarks as realistic, user-driven, multi-turn coding sessions.

Instead of handing over the full task at the start, requirements are surfaced gradually through an interactive user-agent loop.

[Chetan Rane @the_chet2an]

Computer use is having a big moment, and there’s much more to the progress discussion than grindability:

1. Simulation environments are a bottleneck for any RL task, but the software component of building GUI sandboxes isn't the hard part anymore. Implementations like CUA-Gym have already given labs a floor to work with.

2. Computer use has historically struggled with a lack of a good hillclimbing target. Most benchmarks don't simulate non-contrived economic work, and rewards are often brittle and unreliable. OSWorld 2.0 is a strong step in the right direction.

3. We have an interface uncertainty issue. The future is neither GUI-only nor API/MCP/CLI-only. Pure computer use is a painfully slow experience, and we need more hybrid environments and opinionated rewards that discern when tool vs computer use is ideal. Check out what WeaveBench is doing.

@BowenWangNLP @TianbaoX @yuan_mengq43669 have been making some serious contributions to the space

[Mick Gupta @Mickgupta]

Nice. It’s like Perl scripts for non-coders…

[OpenAI Developers @OpenAIDevs]

Show Codex a workflow once. Reuse it as a skill.

Record & Replay lets you show Codex a recurring task, like filing an expense report or submitting a time-off request.

Codex turns that demo into an inspectable, editable skill.

You control when recording starts and stops.

[Alvin Sng @alvinsng]

Imagine being the IT admin who just found out they have to spend their entire weekend building SCIM groups for US citizens vs. non-citizens

[Mick Gupta @Mickgupta]

#truth

[Jason Droege @jdroege]

@GergelyOrosz New leadership here. This isn’t true. Back in the day sometimes people had to roll up their sleeves and pitch in. I’ve never heard any leader say this at Scale and generally haven’t seen any one in a state of horror over anything.

[Aakash Sabharwal @aakashsabharwal]

We continue to get great feedback from model builders on how @ScaleAILabs work is the standard for how to produce and use Rubrics data!

[Scale AI @scale_AI]

This month we turn 10.

The hard work started in 2016, and it hasn’t stopped.

Shortcuts are for losers. Winners welcome.

https://t.co/cPL5Kq8igC

[Scale Labs @ScaleAILabs]

Refactoring is hard, even for frontier agents. SWE Atlas refactors are 2× the size of SWE-Bench Pro and 30× SWE-Bench Verified by lines changed.

[Andrej Karpathy @karpathy]

This is the the quote I've been citing a lot recently.

[Mick Gupta @Mickgupta]

@paulg @bgurley @paulg what made them so impressive?

[Mick Gupta @Mickgupta]

Damn Trisolarians

[Mick Gupta @Mickgupta]

12 dead young scientists --> 3 body problem unfolding in real time...

[Mick Gupta @Mickgupta]

12 dead young scientists --> 3 body problem unfolding in real time...

[Aakash Gupta @aakashgupta]

To get all my takes without an algorithmic filter, subscribe to my newsletter:

https://t.co/vHdRoFmd5z

[Deva Hazarika @devahaz]

Nikita: no more payment for divisive US political ragebait from around the world

Elon: whoa whoa whoa buddy

[Mick Gupta @Mickgupta]

@aakashgupta Not wrong but brutal….

[Mick Gupta @Mickgupta]

I thought the internet would revolutionize buying property and reduce the 6% spend on real estate agents (especially at the higher end). But looks like it’ll take AI to finally get there.

[Ejaaz @cryptopunk7213]

i mean this story is insane.

man used chatgpt to sell his house in 5 DAYS. got 5 offers in 72 hours. no real estate agents. saved so much money doing it too. he used AI to:

> price the house (researched neighboring properties for sale)

> wrote up the legal contracts (saving him $500/hr lawyers)

> best part: MARKETED the fucking property for him (usually you pay estate agents for their network of buyers - ChatGPT did all of this)

i honestly thought sales people would be hard for AI to replace (you need to know people) but apparently not

[Mick Gupta @Mickgupta]

@toddsaunders @bcherny Mind sharing how please?

[Mick Gupta @Mickgupta]

@LongTimeHistory @sandibachom Paid pardon coming soon?

[Mick Gupta @Mickgupta]

@GovPritzker And perjury

[Mick Gupta @Mickgupta]

@GarrettLord Extremely basic take. Thinking from first principles but without any actual insight.

This is closer to the actual truth

[Aakash Gupta @aakashgupta]

This chart is quietly showing you the new playbook for AI coding companies and nobody is talking about it.

Cognition and Cursor both started as wrappers running on Claude and GPT. Now look at this benchmark. Cognition’s SWE-1.6 at 51.7%. Cursor’s Composer-1.5 at 50.8%. Both sitting within striking distance of Claude Opus 4.6 at 53.6% and GPT-5.3-Codex at 56.8%.

Neither company trained a foundation model from scratch. Both took open-source base models and applied reinforcement learning in real coding environments. Cognition’s Swyx said it directly on Hacker News: “it’s increasingly less important the qualities of the base model as long as it’s good enough, because then the RL and post-training takes over and is the entire point of differentiation.”

That’s the thesis. The base model is a commodity. The RL pipeline trained on your specific agent harness, your tool use patterns, your real user sessions is the defensible layer. Cognition trained SWE-1.6 on their Cascade harness with two orders of magnitude more RL compute than SWE-1.5. Cursor trained Composer inside live IDE environments with file editing, semantic search, and terminal commands. Both co-designed the model and the product together.

The math on the jump tells the story. SWE-1.5 scored 40.1%. SWE-1.6 scores 51.7%. Same base model. Same 950 tok/s inference on Cerebras. The entire 11.6 point improvement came from better RL recipes and more compute. That’s a faster rate of improvement than most foundation labs are getting from pre-training scaling.

This is two $10B+ companies (Cognition at $10.2B, Cursor at $29.3B) independently converging on the same conclusion: you don’t need to build GPT-5 to compete with GPT-5 on coding. You need RL at scale on top of a good enough base, co-designed with your agent infrastructure.

The speed layer matters too. Cognition runs at 950 tok/s through Cerebras. Composer runs at 250 tok/s. In agentic workflows where the model loops dozens of times per task, that 4x speed gap compounds into meaningfully different user experiences. Cognition is betting speed plus accuracy beats accuracy alone.

The question that should worry OpenAI and Anthropic: if two startups can get within 5 points of your best models using RL on open-source bases, what happens when the open-source bases get better? Every improvement to Llama or Qwen flows directly into Cognition and Cursor’s pipeline. The foundation labs are essentially subsidizing their own competition.