Come build LLM-apps at University of Oxford. Looking for an @aiDotEngineer to join our team. Especially anyone with experience in AI frameworks like @LangChainAI and evals like @promptfoo or @langfuse and tools like @OpenWebUI Apply even if not perfect fit with the job ad ↓

I am not sure what we've seen so far is indicative of this being a clear conclusion. So far, it seems that there is a certain threshold you need to pass first.

@AnnaRMills The links were to posts by Ethan Mollick... But DR hallucinated a different name while giving accurate quotes.

Link to OpenAI Deep Research on Deep Research: Link to Google Deep Research on Deep Research:

RT @chrisalbon: The best deal on Audible. 280 hours of audiobooks for $10

@perigean @AdamRodmanMD @CoryRohlfsen @DavidDeutschOxf @ToKTeacher What information would calling this "knowledge" or not add? It seems that if we know all the other things, asking if it is "knowledge" is just a different way of asking is it "good". You can ask the same question about an expert diagnostitian's performance.

Interesting accidental jailbreak for @AnthropicAI's Claude when I asked it to give me a TLDR; of Mistral's announcement, it instead summarised its own system prompt. Thoughts why @alexalbert__ and @AmandaAskell?

RT @sama: chains of thought for o3-mini! (we try to organize the raw CoT to make it more readable, and optionally to translate languages,…

Link to application for AI engineer position at University of Oxford:

Talking about the challenge of the every widening knowledge gap leaders face when dealing with AI at #ISE2025 today.

There is a wide range of research tasks where Deep Research doesn't seem to do well. 1. Humanities/social sciences related and 2. Areas where the answers require developing "a feel" for an area before doing the actual research.

RT @karpathy: There's a new kind of coding I call "vibe coding", where you fully give in to the vibes, embrace exponentials, and forget tha…

Battle of "Deep Research": Google vs OpenAI - Same prompt same bad results. I asked for a comprehensive list of Open Source LLMs and both failed at the object level task at making a list of actually current models. ChatGPT Deep research did a better job of finding out information about the models that are no longer current but it took about twice as long.

RT @norabelrose: Our results suggest that the set of features uncovered by an SAE should be viewed as a pragmatically useful decomposition…

Organisations the world around are being urged to "use AI" to increase productivity and redesign their processes for the AI age. But are the companies "making AI" so productive "because they have AI" or because of the boring old formula of "talent + hard work + funding + luck"?

@DonaldClark I agree but I don't know that there is a way "to do it well" yet. The "knowledge gap" is still a problem - it is not immediately obvious how any one individual let alone an institution can do this well. I think the test would be are @OpenAI or @AnthropicAI "doing AI well"?

I'm pretty sure the answer to "does it have goals" is not a straightforward "yes" unless we think "goals" is adding some tokens to previous tokens. We need to differentiate between "goals" and "intentions". The current systems can simulate having "goals" within the process of generating tokens which can be dangerous if connected to tools to enact those simulations but they do not have independent goals with a feedback loop to intentions. "Consciousness" is a woolly concept but in some way it gets at the core of the problem of independent, persistent intentionality better than "goals".

Color me not surprised! At least humans are "truly creative" and AI is just a rehash of what came before was always the worst of all takes on AI.

I kind of feel bad for asking o1 pro to think again on such a trivial problem, but it took 10 more minutes adter a re-prompt and got it right. Close to 15 minutes of combined reasoning to figure out that this spells out "Vertical text makes no sense to LLMs".