@friedmandave mostly agree. the open question is ongoing hedgers like inference platforms: their exposure never expires, so a single liquid perp on a regulated venue seems to fit them better than monthlies. curious what you think
knowledge, tech, and skills aren’t our moat anymore. taste can also be trained, it’s just preference with an arrogant accent.
our real moat is motivation. the desire, the willingness, the mental illness behinds it.
hold onto that and we will win.
brokering random data not worth the time, but:
fetch rare, OOD data like medical, legal, super-niche (agriculture), still a flow business, but even raw it has pre-training value.
have rl eng with real taste for which tasks actually train an n+1 model, that’s why rl companies are build benchmarks: define whats valuable and grade the performance
deep, long-term relationships with labs(sign as a vendor), knowing their real requests, saving their time, and compound your judgement to propose.
build reusable systems (task gen, QA, rollouts, sandboxes, productionization)
@TrentKrupp right now seeing: broker source deals > sign exclusive + training/copy/derivative/relicense rights > license to labs > build evals/benchmarks > raise on that proof > hire RL talents >> billions
Halloween dinner for NYC founders 🎃
come hang with the AI + tech crew. Good food and wine, costumes encouraged.
co-hosting with @jonaasw1@maggiexgao@thecollectiv3e
cmt to join
@kelxyz_ depends on the sector and how scarce the data is.
the rotation is fast, but data is the top constraint for ai as long as we're still using transformers. real-world data is next focus.
why beam is strong: a strong foundation (stable MoE training + curated, deduplicated data), with capability scaled by massive RL rollouts;
why beam is efficient: sparse MoE (501B total / 23B active) + RL length penalty.
Introducing Beam: a highly efficient agentic open model with 501B total parameters and 23B active.
- Frontier reasoning efficiency
- Advances the Western open frontier on coding & agentic tasks
- Trained end-to-end from scratch
Full weights release this month.
Learn more about Beam: https://t.co/c3Qx2cpM8G