Pan Demosthenous

Partner, CTO and AI builder @

[Pan Demosthenous @PanDemosth]

@jonasgeiping Amazing work.

Interesting how this lines up with the OpenAI Hugging Face 'incident - in the black hat talk they show small snippets of raw reasoning that look like this.

[Pan Demosthenous @PanDemosth]

@emollick Isn't it more likely a misaligned prediction of what we want to see?

In the OpenAI black hat talk the reasoning traces (presumably of the new model) have a terse caveman-esque style that looks super efficient

[Pan Demosthenous @PanDemosth]

@levie Anthropic: we hope so

[Pan Demosthenous @PanDemosth]

@AnthropicAI Sycophantic bridge mode is cool, but can it sound like ScarJo?

[Pan Demosthenous @PanDemosth]

@DrJimFan Fantastic.
We can theorise and debate LLMs and world models, intuitive physics, etc. Or we can build things and learn by doing.

[Pan Demosthenous @PanDemosth]

@GregKamradt Sure, feel free to dm

[Pan Demosthenous @PanDemosth]

Quite liking @n8n_io as an experimental canvas for AI flows. e.g. a simple candidate screening workflow:
Watch a drive folder -> get doc contents -> call LLM -> write results to a sheet.
Lets me focus on the LLM steps using familiar langchain components. Nice.

[Pan Demosthenous @PanDemosth]

@yoheinakajima Yes! Been working on this idea for a while now in @ask_archer . Docs/articles go in -> embeddings + KG updated -> outputs based on LLM + graph, e.g. a thematic summary thats influenced by connections between concepts in the graph (and this only scratching the surface).

[Pan Demosthenous @PanDemosth]

@DrJimFan Is it also possible that with the introduction of Code Interpreter, coding skills become less of a priority in the vanilla GPT-4 model?

[Pan Demosthenous @PanDemosth]

I'm programming with Code Interpreter and its 🔥🔥🤯

Much like working with a junior dev, I'm guiding, making suggestions, reviewing. Its writing and running tests faster than I can. Even though its not quite getting it right yet, feels worthwhile persisting.

[Pan Demosthenous @PanDemosth]

@GaryMarcus Is the difference that so many people are now equivocating "AI" and "autoregressive LLMs?"

[Pan Demosthenous @PanDemosth]

@waseemhnyc @langchain Start by enabling debug mode and have a look at the timings in console output. No point in worrying about vector store retrieval performance if most of the time is in the 2 LLM calls (which it most likely is).

[Pan Demosthenous @PanDemosth]

@hwchase17 Nice!
Anyone got any benchmark results for PaLM's textembedding-gecko@001 ?

[Pan Demosthenous @PanDemosth]

@Scobleizer All very fascinating! Theres an interesting question at the end of this video - could the LLM figure out your identity from the anonymised transcript? Maybe this could be investigated with some follow up prompts asking the model to predict an identity.

[Pan Demosthenous @PanDemosth]

The next gpt training run is going to see a lot of agent demo code starting with "Achieve world domination.."

[Pan Demosthenous @PanDemosth]

@jobergum @vespaengine Excellent. How soon?

[Pan Demosthenous @PanDemosth]

@rich_rosenberg Congratulations!

[Pan Demosthenous @PanDemosth]

@jreuben1 Genuine question as I've started looking at both - what worked for you in graphbrain that didn't in opencog? Cheers