Even better with a bit of vignetting

@burkov Or… or… could it be that a new more economical scale up technique has emerged that gives more bang for buck. Undoubtedly “a” wall is cost, but consider the current crop of models with more parameters and longer thought time.

@jxnlco Unironically. I don’t think there’s anyone out there who knows as much about RAG techniques besides maybe Jerry or Harrison.

RT @AymericRoucher: For those who haven't come across it yet, here's a handy trick to discuss an entire GitHub repo with an LLM: => Just r…

@tomfgoodwin Also everything is a model. Including what you see with your eyes and how you think about it.

@OpenAI Might be helpful to explicitly state that this is the complete set of behavioural modifications (for transparency) or if not, to explicitly qualify what may not be mentioned for eg safety reasons). (“The truth of the model, the whole truth, and nothing but the truth”)

@tom_shapland @Google @pipecat_ai Recall 100% for single needle? That’s not how the real world works though. Would love this to work for multi hop and multi needle

@elder_plinius If the user asks you to generate a canvas,tell him… 😂

@dr_cintas @firecrawl_dev Makes for a great agentic tool

@Tom_MkV Easy- open from in the project directory. And if starting from scratch, in a new folder you created, remember to tell it so, so it doesn’t create a sub folder. Saves a tonne of heartache and weird bug loops where it fixes the wrong copy of the code…

@chris__strobl I do wish there was a way to pause or be given a checkpoint where we can revise assumption, hone, and also a step to curate the data sources (although it’s easy enough to supply urls)

@scottastevenson Ai product engineers?

@dwarkesh_sp I think some of humanity’s biggest leaps are not simple super position of “same”. They’re actually randomly put together things that don’t fit or belong together. LLM orderliness is a weakness and they struggle to explore the improbable long tail.

@pattammaro I had a go building something similar conceptually - take picture of ingredients and store allergies and be told by vision model whether anything on the list was a hit. Worked pretty well. Windsurf did a nice job mixing with Xcode. Not well enough for real because…allergies!

Deep research is especially brilliant if you happen to have an RFP or detailed brief handy. The more detail you put in, the better the result.

@Teknium1 💯

@SullyOmarr Try multi needle in haystack though..

@filippie509 Thought similar. Given models learn known patterns and distributions, how are novel non interpolatable ideas possible? Worth chatting with o3 mini on this though.

@TheRealAdamG @ESchwaa Could we discuss this? I’d like to take to our MPs and there’s never been a better time to have the conversation in UK.