gurkanwal singh brar

If Tesla charges $1 per mile and it costs roughly 50 cents per mile, that would result in a gross margin of 50% and likely operating margin of around 35%. Hmm.. what does that sound like? So, if Robotaxis are cheaper and a better experience than owning a car, what’s it worth? Let’s do some quick maths.

@y0b1byte @ShankarAnoushka has been a constant coding companion for a decade now

Best programming music ever. Post yours in the thread.

ਦੁਹਾਈ ਨੀ ਦੁਹਾਈ ਨੈਣਾਂ ਨਾਲ ਹੋਣ ਲੱਗੀ ਨੈਣਾਂ ਦੀ ਲੜਾਈ ਸੰਦਲੀ ਦੁਪੱਟੇ ਉਤੇ ਬੂਟੀਆਂ ਨੇ ਕਾਲੀਆਂ ਕਿਥੇ ਕਿਥੇ ਖਹਿੰਦੀਆਂ ਨੇ ਕੰਨਾਂ ਦੀਆਂ ਵਾਲੀਆਂ ਜ਼ੁਲਫ਼ਾਂ ਦੇ ਕੁੰਡਲੇ 'ਚ ਮੌਤ ਲਟਕਾਈ ਦੁਹਾਈ ਨੀ ਦੁਹਾਈ -- ਨੰਦ ਲਾਲ ਨੂਰਪੁਰੀ

@SullyOmarr vertically integrated agents ftw

Hardest part about building agents isn't even the tech anymore It's balancing "this is an insane demo" vs "this is actually useful" So easy to go down the rabbit hole of building crazy multi-tool planning agents (I'm guilty of this) In practice users don't care. Its just 10x more complicated, fails more, costs more, and has a worse UX

@seema_amble @GoodPoint123 Do you expect agents to fill this gap or copilot like assistants?

@GoodPoint123 yes - i don't disagree - this has been why past process automation cos have had heavy implementation costs/timelines. i do think this changes today with llms

the pyramid below originally showed how finance teams were spending their time towards the bottom on repetitive manual tasks. but! this really applies across a company. a lot of time is spent at the bottom of their pyramid - marketing (e.g. copy, content creation), sales (e.g. outreach), recruiting (e.g. outreach), design (e.g. mock up generation), legal (e.g. contract review), etc. we've seen so many copilots, but really the more interesting opportunity imho is taking some of these boring repetitive tasks - probably picking something very narrow to start - and automating it completely, then going from there

@thatguybg time to see results of your experiments goes up the lower you go into the stack. When you can’t rely on experiments you have to rely on conviction, and nothing strengthens that more than religion

why do hard tech founders seem to be the most religious of any category?

@nealkhosla Totally. Exactly the same argument for Devin / coding copilots and software engineering. A good engineer solves for edge cases, copilots solve for boiler plate which can accelerate but not replace.

a fundamental problem with all of digital health is that it is still feature incomplete relative to the traditional healthcare system both on a micro and macro scale. most of healthcare is about how you handle edge cases and digital health is design to handle the 80%.

@GergelyOrosz definitely one of the few systems where closed loops with formally constructed critics of llm outputs can be built.

Hear me out: What if the single best use case for GenAI / LLM is in the coding domain? Code generation, autocomplete, debugging assist etc. Thanks to coding being one of the very few places where 1. Training data is plentiful 2. Hallucinations can be limited w a feedback loop

@jpdeol do you not like your gas fresh? i see nothing wrong here

@sh_reya I remember seeing some chatter in the crazy days of last year around LLMs carrying a world model. What’s your read on that? Is some level of physics is a required part of a world model?

My mental model of LLMs now is that they are very good at pattern matching, but there are simple operations/rules of the world that they don’t know (provably can’t learn). Kind of like a human that can’t learn some physics. Interesting to build software around these things.

The SWE-Agent repo is the first place I’ve seen the term “agent-computer interface” (ACI). I find it interesting that 2/4 ACI features have to do with cutting irrelevant context. The other ACI insights imply that LLMs don’t reason about program outputs well

@officialKrishD @morgancheatham yep makes sense, seems like note taking will be a feature into existing products much like google meets, zoom adding fireflies like note taking into themselves

In my opinion, it went into a spiral. First SUKI ( if am not wrong ) created a clinical note taking assistant for doctors. Then Nuance Communications( acquired by Microsoft mostly ) tried to launch an AI note taking service. Think AWS also launched a services under AWS ComprehendMedical ( can’t recall but they discussed about note taking through AI and then codify to medical standards) Then very recently , another startup made quite a noise around Note taking feature from zoom ( especially clinical , can’t actually recall the name ), but ended up bringing it to open source mostly. The speech synthesis projects were never meant to be a multi million dollars deal; when the emerging language generation through generative models are getting better day by day.

what's a category in healthcare that venture capital was bullish on, but got wrong? and why?

@anshublog Wallet holders make stuff happen :)

Things you don't expect till you actually see it. You send 13 emails to 5 directors and 8 VPs at a trillion dollar software company: crickets. You send 1 email to the CEO: yes, let's partner. x will follow up.

@medjournalist @morgancheatham Do any of them have traction?

@brargk @morgancheatham A gaggle of these now

@adamsilverman closer to 99 i’d say. But from what i can tell, agent is a nice gtm strategy - as they onboard customers they will find the exact vertical use case the agent is needed for and build guardrails to serve those. Effectively ending up as a ML Api company, which is not a bad end goal

Hot take: 95% of “agent” companies lack the reliability required for widespread consumer or enterprise adoption.

@sudobunni because some teams start treating it as the one true god, as against another tool around which the processes should change as the team grows

Why does everyone hate Jira and Agile so much?

Yogesh and I are excited to announce we have raised a $2.8M pre-seed round! The last 8 months have been an amazing adventure, but there’s so much more to come. We are expanding the team, building AI training infra that’s never been built before, and launching our protocol 👇

@max_paperclips @rao2z @sh_reya I agree, but in cases where a formal method isn’t available and you need human verifiers , you need to gradually eat up their verification hours with a model.

@brargk @rao2z @sh_reya Well, ideally it's not a "verification dataset" necessarily - it's an external tool such as a code interpreter or theorem prover. If it's generating knowledge, ideally you have an ontology of some sort that you can validate the information against to whatever extent possible

Getting reliable outputs from LLMs is tricky. “Use LLMs to validate LLM outputs,” they say, but who validates the validators??? My collaborators & I are interested in feedback on our ideas to solve this. If you have a spare hour in the next few days, please DM/email me 🙏 thx!

@rao2z @sh_reya The key is identifying when an input - output pair is novel, or out of distribution for the verification dataset . I haven’t hit this yet, but iam hoping conventional confidence based approaches can help here. @rao2z do you have any pointers on this front?

@rao2z @sh_reya Absolutely, a slow but compounding approach is introducing humans in the loop to criticize / verify and gradually building out conventional models to replicate them. You still need humans in the loop even as verification dataset grows because novel scenarios arise.