You walk into your nearest Chase (or similar) branch. Out in 10 mins.
964 karma · joined September 19, 2017
Practical, real-world application.
I got to present at GCP Next about a part of this last year: https://www.youtube.com/watch?v=5QsM1K9ahtw
I’m presenting in one (and maybe two) sessions with more info on the training side this year.
There is a reason Cursor is the fastest startup to $100M in revenue, ever.
Claude Code, Cline, Cursor… all of them with Claude 3.7.
The results of this paper would indicate doing what they did, but online could return better results
Receipt scanning is a multiple orders of magnitude more valuable business. Mistral at this point is looking for a commercial niche (like how Claude is aiming at software development)
Source: We have large EU customers.
Not a single example on that page is a Purchase Order, Invoice etc. Not a single example shown is relevant to industry at scale.
INT8, INT4, FP8 and soon FP4
If GPT 4.5 had nothing of value to offer other than a flex of opening eyes ability to scale, it’s still a signal that they have the chops to throw the most amount of compute at the upcoming reinforcement learning race.
If you’re actually selling enterprise solutions in this space, you’ll quickly learn that a large number of Enterprises have their own private deployment of open AI on Azure and are pushing their vendors to use that even if it means they get lower quality outputs for certain use cases.
Data and model quality aren’t the only moat-able features.
“A society grows great when old men plant trees in whose shade they know they shall never sit”
Take a set of prompts, run it through the large model and small model, calculate KL divergence between all the logits, then update the small model to minimize the loss.
This gives the smaller model a much higher density signal than SFT which is usually cross entropy against the single correct logit.
Edit: note that the term is abused quite often and SFT on just the traces from the big model is also sometimes referred to as distillation (Eg the Deepseek smaller models). IMO that is incorrect.
When you combine that fact with being part of a diverse batch of requests over an MoE model, outputs are non-deterministic.
This still puts domestic sites at a significant disadvantage.
Deepseek showing that you can do pure online RL for LLMs means we now have a clear path to just keep throwing more compute at the problem! If anything we made the whole "we are hitting a data wall" problem even smaller.
Additionally, its yet another proof point that scaling inference compute is a way forward. Models that think for hours or days are the future.
As we move further into the regime of long sequence inference, compute scales by the square of the sequence length.
The lesson here was not "training is going to be cheaper than we thought". It's "we must construct additional pylons uhhh _PUs"
Markets remain irrational and all that...
His Majesty’s government is struggling to keep its people warm and fed, let alone show any economic growth.[0][1]
The UK needs more foreign investment than ever before.
[0] https://www.bbc.com/news/uk-68625344 [1] https://archive.is/2024.12.15-052556/https://www.ft.com/cont...
The law makes it so.
They were a key cause for LLMs being a thing in the first place.