Kevin Hu

🏘️

[Kevin Hu @_newhaiku]

probably nothing

[Kevin Hu @_newhaiku]

three years ago we started building tools for agents that didn't exist yet

they exist now

[John Kim @john_sungjin]

Today, we're launching Village to let investors orchestrate teams of agents to scale their judgment.

Three years ago, we bet that both humans and agents would need new tools to fulfill the promise of LLMs to transform research. Village is that tool, the first IDE for research built from the ground up so your agents can answer bigger questions, learn from your feedback, and build the AI research infrastructure for your firm.

We're starting with public equity research, enabled by our EDGAR and earnings call transcript datasets, and we'd love to hear what use-cases you'd like to see next.

The feeling of manipulating huge amounts of information and collaborating with hundreds of Assistants is really special, and I'm excited for you to experience it. Our beta sign up link is below!

[Kevin Hu @_newhaiku]

@A_Reichenbach_ congrats on the launch!

[John Kim @john_sungjin]

making our @conviction embed submission public to share some long-overdue writing on:

- why vertical AI products haven’t hit widespread adoption among investors
- our differentiated product vision
- why we’ll win

link below. lmk if you love it or hate it

[Kevin Hu @_newhaiku]

@andersonbcdefg one of the worst AI SDKs in the history of AI SDKs, maybe ever

[Kevin Hu @_newhaiku]

gemini refuses to talk about openai and chatgpt–and even questions about anthropic get cut off mid-response

[Kevin Hu @_newhaiku]

@runpod @ryancruzb check your dms!

[Kevin Hu @_newhaiku]

@JesseDodge Why is OLMo 7B the Pareto frontier when Mistral-7B achieves 62.5% five-shot MMLU? https://t.co/NT26zzN4nP

[Kevin Hu @_newhaiku]

@random_walker @simonw That's incorrect. There's a 32k tokens/min limit as well as a hard cap on the number of output tokens. See https://t.co/Z7XKldI0Ie

[Kevin Hu @_newhaiku]

@nlsanders @abacaj nvitop --colorful

[Kevin Hu @_newhaiku]

@abacaj .from_pretrained(..., device_map=device) instead of calling .to(device) afterwards should speed things up

[Kevin Hu @_newhaiku]

cancel culture has gone too far

[Kevin Hu @_newhaiku]

@jerryjliu0 @disiok Great guide on fine-tuning! The evals could use some work tho:

- BGE works best with instructions, but these aren't prepended.
- Half the texts exceed BGE's 512 token limit and are silently truncated.

After adding both, vanilla bge-small-en's hit rate is 0.87, on par with Ada.

[Tejal Patwardhan @tejalpatwardhan]

We’re excited to write about our experience using LangSmith to simulate multi-agent social networks! @langchain https://t.co/yHpiEu7Jbw

[John Kim @john_sungjin]

We got 🥈 at the @AnthropicAI hackathon by impersonating our favorite anons and judges -

but we're really excited about the idea of using LLMs to Simulate Everything! So, we created Twitter (we can legally call it that now @elonmusk ).
1/n

(❤️ @kvvnhu @taehyoungjo @thetejmahal)

[Aman Sanger @amanrsanger]

gpt-3.5-turbo is criminally underrated at coding

When using it with Azure's completion endpoint instead of OpenAI's chat endpoint, you can get a jump in HumanEval performance from <50% to 74%!

This blows claude v1.3 out of the water, which sits just below 60% perf. [1]

[Kevin Hu @_newhaiku]

"—Drastically increases the risk of flameout when the private markets correct." You got that from a Bill Gurley interview, right? Do you have any thoughts of your own or were you just gonna plagiarize his whole argument about startup overvaluation and the coming correction?

[Kevin Hu @_newhaiku]

Of course that's your contention. You just launched your first company after reading The Lean Startup—you're convinced that constant iteration and pivoting based on customer feedback is the only way to build an empire.

[Kevin Hu @_newhaiku]

—Well as a matter of fact I won't, because raising huge rounds at sky-high valuations drastically increases the risk of flameout—