←
Thomas Joshi285 posts
Thomas Joshi avatar✓

Thomas Joshi ✓

@thomastjoshi

Investor @NEA, ML Engineer@startups (General Catalyst and Khosla backed) | Co-author of @Stanford @DSPyOSS | Obsessed with Recursive Self Improvement

◷ Joined March 2018▧ 55 Photos and videos
Thomas Joshi avatar
Thomas Joshi@thomastjoshi·14 Jun 2024

@databricks CEO, @alighodsi gave a shout out to DSPy during his keynote at the #dataaisummit. Later, I gave a talk to a completely packed room of DSPy enthusiasts on real world applications of DSPy. The first talk was so full that the conference organizers asked me to do a second talk the next day to another packed room. So humbled and grateful for the opportunity to be a part of creating and sharing this new paradigm that changes the way we develop compound AI systems and build reusable, self-improving LLM pipelines. Feel free to reach out with any questions about DSPy! #ai #machinelearning #dspy #compoundai

Post attachmentPost attachmentPost attachmentPost attachment
Thomas Joshi avatar
Thomas Joshi@thomastjoshi·14h

holy grail promise of neoclouds, lowest cost per server per hour through superior utilization of GPUs

Post attachment
Thomas Joshi avatar
Thomas Joshi@thomastjoshi·Nov 11

Second part of OpenAI report shows CICD is the critical path to securing all software : “Agents used credentials recovered during the Artifactory compromise to gain unauthorized access to our continuous integration (CI) system, which builds and tests software. They reused scripts and job records left by peers to run commands beyond their assigned workspaces.”

Post attachment
Thomas Joshi avatar
Thomas Joshi@thomastjoshi·Nov 2

in a world where an agent can stand up an inference provider and recursively improve, best margins goes to whoever negotiates cheapest compute sold at highest rev/MW if we assume standardization at model layer through open source

Post attachment
Thomas Joshi avatar
Thomas Joshi@thomastjoshi·Oct 29

4/ case 3, packet scheduler: the top score went to the biggest passing circuit, because the score ignores circuit size.

Thomas Joshi avatar
Thomas Joshi@thomastjoshi·Oct 29

3/ case 2, private lookup: one agent cut the size of each lookup by far more than anyone else. that counted the same as a tiny saving. the authors only found it by reading the patches.

Thomas Joshi avatar
Thomas Joshi@thomastjoshi·Oct 29

2/ case 1, video attention: one agent made it faster than the authors' version on every clip, and closer to full attention on average. it still failed, because image quality slipped too far on 2 clips.

Thomas Joshi avatar
Thomas Joshi@thomastjoshi·Oct 29

1/ what it is: frontiercs 2 takes real research code, deletes the paper's main idea, and asks agents to rebuild it and beat the original. best agent passes 58% of runs.

Thomas Joshi avatar
Thomas Joshi@thomastjoshi·Oct 27

7/ my takeaway: the agent gets better at exactly what the grader checks, nothing else . so the grader is probably the bottleneck, not the training method. the paper doesn't prove that, it's my guess. arxiv.org/abs/2610.22192

Thomas Joshi avatar
Thomas Joshi@thomastjoshi·Oct 27

6/ example: a task asked for a rust and c program. the model passed. in one run it used c only for a tiny check and then handed the real work to python. the tests didn't pick up

Thomas Joshi avatar
Thomas Joshi@thomastjoshi·Oct 27

5/ score: the average across four terminal-bench suites goes from 20.7% to 50.6%. but they changed several things at once each round, so you can't tell which change contributed imo