Chris Poensgen

founder, building a lab for continual learning for enterprise

[Chris Poensgen @Poensgi]

@psantacl anthropic asked for this

[Chris Poensgen @Poensgi]

@KanishkaNarayan @AnthropicAI not just the base models need to be sovereign, even the vertical specialized models have to be owned by enterprises to stay independent

[Chris Poensgen @Poensgi]

@aidangomez we need a marketplace of frontier labs from different schools of thought

[Chris Poensgen @Poensgi]

@BlackHC 1. build compute yesterday 2. get the best people in a frontier ai lab filthy rich building a competitive open source model 3. subsidise enterprises post training specialized models for their use cases on top of that model.

[Chris Poensgen @Poensgi]

@DadaJudith it’s one thing to depend on m365 subscriptions, another if the ceiling of intelligence gets defined outside of europes control

[Chris Poensgen @Poensgi]

@mweinbach watch out your limits are going to be blown in a minute

[Chris Poensgen @Poensgi]

hey @HarryStebbings loved the discussion about oss models but i think you should take this discussion a few years forward. when looking at nvidias just released chips you can see the industry is preparing for a future where in just a few years current level intelligence will be on local hardware. not sure that future makes me incredibly bullish on the frontier labs.

pcs went from the size of a house to a slap in your pocket before and we will see it again!

[Chris Poensgen @Poensgi]

@tibo_maker ask @arvin_haddad

[Chris Poensgen @Poensgi]

@DadaJudith @nvidia @kyosu @mpvl_ @bfl_ml @tobiashalloran thank you, the event was great!

[Chris Poensgen @Poensgi]

@miramurati @andrewarruda going to try this out in behaviour simulations. with multiple agents interacting in a 3d world turn-based falls apart quickly

[Chris Poensgen @Poensgi]

@DadaJudith @nvidia @kyosu @mpvl_ 🖐️

[Chris Poensgen @Poensgi]

Code: https://t.co/anQYHeXvSZ

[Chris Poensgen @Poensgi]

Emotion-steered LLMs at 300+ tok/s on an L4 with batching.

We built open source library that lets you serve an OpenAI-compatible API where request pick its own affective direction from an api parameter, fearful, surprised, angry, without breaking vLLM's batching.

Open source, with a write-up of how it works:

https://t.co/Dz6Ei3sO2t

[Chris Poensgen @Poensgi]

@awnihannun @computational Clean.

[Chris Poensgen @Poensgi]

@DennisonBertram @AnthropicAI @bcherny have seen this multiple times 🙃

[Chris Poensgen @Poensgi]

@brycent @matt_kolm also future consumer use cases might be 100x to expensive today, so people can’t fathom them

[Chris Poensgen @Poensgi]

@lucasmeijer LLM-Personalabteilung

[Chris Poensgen @Poensgi]

@craigzLiszt who’s gonna fix it when ai breaks - ai will ofc

[Chris Poensgen @Poensgi]

tip for @karpathy llm-wiki: instruct claude to use pdfgrep for the raw folder.

[Chris Poensgen @Poensgi]

@levelsio to be fair the leaked product in these screenshots is supposed to enable more people to build for smaller audiences. i think the 100k/y to 5m/y businesses are likely fine