Jack Hau

Impressive from @WecoAI! So what’s next… Autoresearch³?

@botirkhaltaevv Surely the multiplier is a function of how well you use AI and it is not the same the how technical you are. There might be some correlation but I wouldn’t rule out the (non-technical) meticulous context gatherer over a (technical) 1-shot wonder prompter

@jackhau0212 AI is a multiplier the more technical you are the larger you're coefficient.

AI has widened the gap between the technical and the non technical. Not the opposite.

I’ve been using @harborframework for a while for evals, and I’m a huge fan. I love the design and the core principles behind how an environment is modelled. The thing that gets me most excited is that good evals unlock so much more than benchmarking. They become the foundation for hill-climbing, auto-research, RL, GEPA, trajectory analysis, SFT data generation, and more. Rollouts, rollouts, rollouts.

OPSD is kinda OP

@bclavie @aiDotEngineer would love to chat to you and learn more about retrieval

very excited about the upcoming @aiDotEngineer is going to be very nice, looking forward to meeting a ton of people I only really know through their profile picture

@alex__mackenzie would love to join 👀

Hosting another eng gathering in London this month (~20 places), come along (just DM) if you enjoy discussing: Distributed Systems Databases High Frequency Trading Web Browsers Compilers AI Infra ++

Excited to release the Ultimate guide to RL environments! Definitions of RL environments differ wildly in the LLM era, so we spent the last month building several RL environments across 6 different frameworks, domains and complexities to map out which are easiest to build with and which can be scaled to 1000s.

@Vtrivedy10 genuinely love the framing of fitting an agent and this has given me a lot of inspiration (thanks!) do you not think that a better mental model for this would be agent is harness(model), eval is the loss function.

Data Driven Agent Design with Evals & Hill Climbing Algorithms this is a mental model dump i’ve been thinking through + iterating on as we’re building self-improvement infra around agents: - mining Trace Data to find errors and tweak the agent harness - building + maintaining evals - using evals to guide the agent update/generation process What are the inputs for fitting an agent: the main idea is what does a sklearn fit(model, data) function look like for agents agent = fit (model, harness, evals) Data Driven Agent Design: Evals are training data for agents - every eval encodes behavior we want to see in our agent, just as every training data point in standard ML produces a gradient to shift model weights Every eval we fit our agent towards votes for how to alter the harness to make that eval pass “Can I start from an empty harness. Fit to evals, and produce a great agent? Should that be how we do things + inject some more human priors” Verifiable Signals: Evals are akin to specs but better because they’re verifiable and measurable We can use rubrics that give a dense feedback signal for work, programmatic verification, and LLM as a judge to evaluate subjective behavior You can directly measure which evals pass and which don’t at a glance. This is very useful, you can’t as easily do the same attribution via a simple markdown spec Iterating on Evals over Time: Evals are model and harness dependent. “Spring Cleaning” and updating of Evals is important as you no longer need them Agent design is focused on vibes today, but could benefit by being open to more data driven design. Just as frontier labs spend millions on data quality, teams can invest time into fantastic eval curation and design. The future of specialized agents will depend on encoding that specialization is something measurable. Data matters, and good evals are a data signal to build a good agent 🚀

Pretty cool stuff: Vector search straight from Parquet files using DataFusion. No need for fancy vector DBs - just tune page sizes to match embeddings - embed an IVF index in metadata - and query via SQL. 5.7x speedup with high recall.

Welcoming Claude, @AnthropicAI's frontier AI model, as the team’s Official Thinking Partner! Through this partnership, Claude will be integrated across the entire Williams organisation—working alongside engineers and team strategists to support how the team thinks, plans, and performs. Read more about the partnership – and what it means for our mission to get back to the front of the grid - here: bit.ly/46sYJtg

SpaceX has acquired xAI, forming one of the most ambitious, vertically integrated innovation engines on (and off) Earth → #xai-joins-spacex" target="_blank" rel="nofollow noopener">spacex.com/updates#xai-jo…

is model pruning just cardio for neural networks?

lol, not sure how I feel about this… is it this legit or google propaganda?

👀

LLM memory is considered one of the hardest problems in AI. All we have today are endless hacks and workarounds. But the root solution has always been right in front of us. Next-token prediction is already an effective compressor. We don’t need a radical new architecture. The missing piece is to continue training the model at test-time, using context as training data. Our full release of End-to-End Test-Time Training (TTT-E2E) with @NVIDIAAI, @AsteraInstitute, and @StanfordAILab is now available. Blog: nvda.ws/4syfyMN Arxiv: arxiv.org/abs/2512.23675 This has been over a year in the making with @arnuvtandon and an incredible team.

filesystem is all you need for context management

TIL a hexahectaenneacontakaiheptagon is a polygon with 697 sides

TIL Colorado has 697 sides