AI in 2024 - efficiency over model size (Nick Jakobi)

[@lexer_]

This is a product sales person, not an ai person.

[@InfiniteQuest86]

Scary. Someone who admits you have to really contort yourself with prompts to get accurate output is also saying he doesn't know if his kid even needs school. There's a real cognitive dissonance going on there. Until we figure out how to make these things accurate (probably an extremely hard problem given the amount of funding already put into this topic), everyone needs to become as intelligent as possible.

[@byrnemeister2008]

The Cohere team just seem a lot more grounded in the real world than the others. To much Silicon Valley BS and main character syndrome. But the Cohere team are probably good folks to go to a bar with.

[@thisisashan]

2024. Humans still saying 'first' and 'second'. AI already won the great robot war.

[@RickeyBowers]

Plot twist - they are robots communicating with other robots.

[@prithwiparthadasgupta4869]

​@@RickeyBowersas one of them i completely agree.

[@DelandaBaudLacanian]

Have you interviewed anyone who is attempting to (philosophically/theoretically) bridge this AI agency gap you mention at 32:50? I think we will see models gain agency in the very near future (despite what the "stochastic parrots" enjoyers say). Great discussion and production as always!

[@alexandermoody1946]

The question that should be highly considered is. Are a series of 8 billion parameter models that are structured using a higher refinement level to the data going to be able to be more effective when weighted alongside a highly generalist model. The Internet was not designed for machine learning. With design of very specific data sets for narrow field relevant understanding, are smaller models coupled to a generalist model able to encourage the generalist to specify appropriate actions for searching for appropriate answers then used for a multitude of scenarios rather than a really large model running higher resource costs?

[@dankprole7884]

Someone needs a punctuation model 😂.

[@alexandermoody1946]

@@dankprole7884 sorry please accept my apology, I wrote that quite quickly whilst working the other day. Edited above.

[@dankprole7884]

@@alexandermoody1946 sorry I am a bit of a troll sometimes! I agree with you - we need to stop searching for the perfect AGI and work with what we have. Personally I will use LLMs only for individual function calls and take care of the logic myself, until I am confident a model can do it for me. That time still feels very far off to me. But for use cases with a ton of data e.g. customer service, i don't think the decisions or logic are that hard or risky, and can be done with agents today. And now we have access to a whole new class of (admittedly buggy) language functions that were prohibitively expensive before, since they required training a specialized NLP model.

[@alexandermoody1946]

@@dankprole7884 I am grateful, I would of forgotten to check back on what I wrote If you hadn't commented and I am much happier with how my initial comment reads now. I do not tend to joke in the comments section so I can come across as critical for many of my comments. In the workshop where I work we play on words, make the worst dad style jokes that are so silly and pretty shameful. That kind of nuance I am concerned may become lost in translation with large language models. I believe we are quite lucky to have arrived in a position where we have the possibility for real enjoyment to share with each other but how we translate that understanding in forms of data may require consideration from a data consolidation and cross meaning architecture that allows context that we may inherently understand.

[@TBOBrightonandHove]

Enjoyed Nick Jacoby's perspective at the cutting edge of making useful AI products for Enterprise.

[@geertdepuydt2683]

You're not on LinkedIn, go away

[@alia.4524]

aya-expanse-32b has to be one of the worst models I've seen on lmarena, I've never once chose it. I'd say out of 30 or so times, maybe more. I look it up and it's made by cohere. Not a good look, acts like a 3b model

[@prithwiparthadasgupta4869]

Second

[@LostInTheRush]

first