Shrey Modi

@forwarddeploy would love to join!

Thinking of hosting more chai and samosas in SF at ours ☕️ Want to come hang out with good people and have fun snacks, lmk below! 🙌

@bingxu_ will the code be open-sourced?

This may be one of the first real signs of superhuman intelligence in software. On some of the most optimized attention workloads, agents can now outperform almost all human GPU experts by searching continuously for 7 days with no human intervention inside the optimization loop. Terry and I started agentic coding efforts at NVIDIA 1.5 years ago. Neither of us knew GPU programming, so from day one we pushed toward fully automated, human-out-of-the-loop systems. We call it blind coding. Over those 1.5 years, the two of us generated 4 generations across 2 agent systems. Since the 2nd generation, the stacks have been self-evolving. Each agent is now around 100k non-empty LOC. When we released the blind-coding framework VibeTensor in January, the implication was easy to miss. AVO makes the signal clearer. My bet is: blind coding is the future of software engineering. Human cognition is the bottleneck.

@LakshyAAAgrawal @gepa_ai Great work!!

Excited to release @gepa_ai's optimize_anything: a universal API for optimizing any text parameter. It consistently matches or outperforms domain-specific tools optimizing code, prompts, agent harnesses, cloud policies, even visuals! If you can measure it, you can optimize it.

@VihaanJagiasi @cadastral_ai Let's fucking go!!

lil life update: left waterloo last september and moved to nyc to join @cadastral_ai as a founding engineer. every CRE analyst in new york is manually reading 200-page leases and underwriting deals in excel. trillions in transactions. the entire back office is still done by hand. we're fixing this. we build ai agents that automate commercial real estate end to end — lease abstraction, due diligence, underwriting, compliance. we already work with most of the big CRE firms in nyc. JLL, Blackstone, AvalonBay, Equity Residential, Empire Realty Trust, Continental Realty etc came out of stealth last week and announced our $9.5M seed round. we’re growing the team and hiring more engineers. in-person nyc.

Convert production data into evals through semantic clustering!

@katieruthmishra would love to attend!

new year, new game! playing soon and this founder will teach everyone how to play PLO soooo you don’t need to know *how to play*. but you must be a degen. comment / DM if you are said degen

for teaching continual learning and memory management, can we do cross episode training where apart from giving the answer at the end of the episode, the model also writes to a file which it can then retrieve later things that it has learnt from the particular episode? the model will get rewards for relevant memory used in future episodes, this way it can learn how to write important memory?

@dakshgup great company to work for!

we're looking for great engineers at greptile to build agents that catch bugs. over the last year, we've grown from zero to millions in revenue, thousands of customers including top engineering teams of all sizes like scale, brex, whoop, substack, partiful and multiple F500s. software is eating the world, and autonomous validation is an extremely difficult problem. if we succeed, there is an opportunity here to build a large and important company. this is a particularly interesting time to join the team. we have laid the groundwork to be able to work on a variety of hard and interesting problems. i'm especially proud of the early team we have built. it's hard to overstate the joy of getting to work with brilliant people that care deeply about their work and about each other. if you're excited about a future that is free of bugs, we'd love to hear from you :)

@lateinteraction will training models specifically for using the RLM harness better create further improvements? curious to know!

Something we perhaps understate in the paper/thread is that *open models can already work extremely well as RLMs*. I was rather surprised by how well Qwen3 does.

@JayadityaSethi lmao

Santa got me what I wanted

maybe even train both the agents together as a system

we should then RL the main agent to initially explore more in the early episodes so we can collect a lot of feedback from the environment and then exploit in later episodes

Wondering if coding agents understand huge codebases mostly because they’re great at context narrowing (e.g., smart grep/search). If so, can we train a similarly strong “context-narrowing” sub-agent for other tool-call / text2sql tasks, one that selects the right domain knowledge + past mistakes to pull in each run. we would need to find a nice memory structure for storing these past mistakes

@YulunJiang @LiangzeJ @DamienTeney @Michael_D_Moor @mariabrbic Find the paper here - arxiv.org/pdf/2512.16848

Really interesting work! This is what I have been thinking for a while, wondering how something like GEPA with this would work Authors: @YulunJiang @LiangzeJ @DamienTeney @Michael_D_Moor @mariabrbic

@dakshgup lets go!!

we just opened a research engineering role in san francisco building the universal bug catcher comes with a lot of interesting and hard subproblems. - how can agents learn the idiosyncrasies of a codebase as fast as possible? - how can agents adapt to user preference to only surface the type of code feedback that they would care about? - exactly what synthetic tests should agents generate and run to detect missed edge cases? if these sound interesting to you, you should reach out to us. we are particular excited about new grads (including masters and phd programs) that have worked on coding agents before or generally interested in applying LMs to software production. this is in-person in san francisco only

Kudos to @lateinteraction @LakshyAAAgrawal for the @DSPyOSS GEPA integration!

We did a few experiments running GEPA and GRPO together to push the performance of open-source models even further! From our experiments you can squeeze out ~60% of the performance using just prompt optimization and then ~40% using GRPO :)