Florian Findeis

Building in industrial AI

[Florian Findeis @Flowrian135]

https://t.co/jFzXvEnwng

[Florian Findeis @Flowrian135]

@Rivian_DD @bigern_ern What is your prediction for Q3 automotive revenue? Guess we have a proxy for Thursday in 2 weeks then

[Florian Findeis @Flowrian135]

@bigern_ern @Rivian_DD I presume you base your assumption purely on VIN. If you look actual registration (prolly the fairest overlap to generally sound accounting) you clearly see the ramp but absolute numbers and revenue recognition will largely ramp in Q4

[Florian Findeis @Flowrian135]

@arankomatsuzaki So agents have already mastered the most human part of research: claiming more than they did 🤣

[Florian Findeis @Flowrian135]

@RampLabs @NVIDIAAI Only 10k examples. How much of the win was label consistency rather than volume?

[Florian Findeis @Flowrian135]

@ApolloResearch First model to worry about its performance review.

[Florian Findeis @Flowrian135]

@awiltschko The valleys aren’t data-poor, they’re data-private. A train maker has decades of drawings, change orders and test logs. Almost none of it ever left the building.

[Florian Findeis @Flowrian135]

@alexandr_wang holy steve jobs 🙃

[CERNpress @CERNpress]

[Press Update]: CERN and Mistral join forces to accelerate AI for science

Find out more: https://t.co/mNlSWcnMeZ

[Florian Findeis @Flowrian135]

@Jsevillamol Experts judge the process, customers judge the result. AI wins wherever only the result gets checked.

[Florian Findeis @Flowrian135]

@KSimback My bet: the boring middle. Back-office history from non-tech firms: quotes, orders, disputes. Least glamorous, least scraped, most of the economy.

[Florian Findeis @Flowrian135]

@BrendanFoody Which job gets its env last? My bet: anything where you only know it was done well six months later.

[Florian Findeis @Flowrian135]

@jonah_b Depends where the supply comes from. In Germany ~114k small firms a year plan to close because the owner retires and nobody takes over (KfW). Those aren’t failing businesses. That’s decades of working processes.

[Florian Findeis @Flowrian135]

@TheyCallMeChew @troyaitken_ Agree on trust. With that few buyers and different mandates, the real work is matching: finding the three whose mandate fits this exact data. Blasting it to everyone burns both sides.

[Florian Findeis @Flowrian135]

@gokulr imo the valuable part of those traces isn’t the code, it’s every not like that from the user. once more matter of taste

[Florian Findeis @Flowrian135]

@adithya_s_k @scaleai @earakely @scale_AI Which parts come from real data: arrivals and weather? The shipping-line agents look like the hardest part to make realistic. How a real planner and a line actually trade off a delay lives in their emails, not in public port data.

[Florian Findeis @Flowrian135]

@OfirPress Outside software, #2 and #3 rarely overlap. Most tasks people actually solved never hit the web: a supplier quote, a design change, a failed inspection. They sit in company archives with the outcome attached, which also keeps them out of training data.

[Florian Findeis @Flowrian135]

@stefkrsteski If you go with MAPS, use its GAIA-v2-LILT set: tasks adapted per language, not just translated

[Florian Findeis @Flowrian135]

@willdepue imo three kinds of data bottleneck:
- data you can generate (code, math)
- data you have to wait for (bio, astronomy)
- data that already exists but is locked away (every company’s quotes, orders, inspection reports)

last one feels like just a permission/access problem