Harsh

robotics research, wams, representation learning, etc. prev @usc

[Harsh @harshwithit]

The “sf has no taste” discourse is overblown and is simply an exercise in creative writing for those who partake

[Harsh @harshwithit]

@jasminewsun For those in the trades - what were they working on before, did they transition careers due to demand? Also would be worth examining the labor components of maintaining data centers after buildout.

[Harsh @harshwithit]

@pdrmnvd then you repeat this after checking that the lmsys leaderboard has a new pareto optimal for the embedding or llm model

[Harsh @harshwithit]

@micsolana hi hat is awesome since they also sell by the slice

[Harsh @harshwithit]

@vikhyatk @justalexoki on multimodality Gemini is far superior in real world usage than any of its competitors

[Harsh @harshwithit]

@mattparlmer Yeah, MMMU is a good proxy for spatial reasoning across models that support multimodal inputs. Gemini Pro/Flash remains at the frontier by a wide margin for performance as well as latency in my experience

[Harsh @harshwithit]

@ericjang11 everything is computer

[Harsh @harshwithit]

@skooookum the benefit is that via in-context learning, once such a practice is established initially, agents pop off

it's a bit like having a jr. engineer onboard a greenfield vs mature codebase -- in the latter, there are way more best practices to mimic

[Harsh @harshwithit]

@OfficialLoganK Remove rate limits on logprobs!

[Harsh @harshwithit]

translation is all you need

[Harsh @harshwithit]

@imranchaudhri congrats on shipping

new schools of thought should always be welcomed with an open mind and in this case it will be interesting to see if widespread usage results in a killer workflow

[Harsh @harshwithit]

@qdrant_engine + @modal_labs is op

[Harsh @harshwithit]

@tszzl is the generalized interface for AGI in humanoid form?

[Harsh @harshwithit]

1. Data companies - Scale AI, Twitter/Reddit (API changes are a big bet on this), Databricks
2. Compute: GPU companies, cloud compute companies, and companies that train foundational models (we know the players) here
3. Interface - ChatGPT, Poe, Tesla Car/Bot, Copilot, TikTok

[Harsh @harshwithit]

data, compute, or interface is how I categorize AI companies

[Harsh @harshwithit]

@chipro On control: Relying on APIs bets that the particular model or company will always be SOTA - using open source allows folks to plug and play as things progress and iteration might be better with OSS even though initial setup is trickier than an API call.

[Harsh @harshwithit]

@zswitten "everything should be deterministic" mentality

[Harsh @harshwithit]

@leahculver they lost a legend... onwards!

[Harsh @harshwithit]

Tbf I can’t remember the last thing I used that felt “timeless” so could be a spurious correlation I’m drawing on here

[Harsh @harshwithit]

Sometimes it’s easy to forget that most target markets are outside the sphere of hyperactive influencers - might be the reason why some products show a growth initially and then stall within 2mo of launch. Echo chambers produce seasonal products