Published some thoughts on how we may be able to unlock “AI coworkers”.
This will be a new class of AI native product across domains where precision is of paramount importance and correctness can be quantified.
I'm excited about the startups building
Bumped into the man himself, Jensen, and convinced him to spend a few minutes with a group we were hosting at dinner.
He showed up and spent 40 minutes. Warmest, most humble and authentic leader I've ever met. And technically so deep across a broad range of topics!
The truly valuable unlock for most enterprises will be when deep learning significantly outperforms tree ensembles on tabular data, such that the deployment costs are worth it.
h/t
@rasbt
for providing a comprehensive overview of developments in this area
Amen to this. Two reasons:
1. Palantir did a good job of screening for raw material
2. Technical people were forced to be super customer centric and develop tons of resilience in the face of tight deadlines and tons of firefighting.
It's a privilege to work alongside incredible colleagues
@kleinerperkins
as we continue to serve the intrepid founders breaking new ground to build a better tomorrow. Onwards and upwards!
Today
@kleinerperkins
is announcing KP21 an $825M fund to back early stage companies and KP Select III, a $1.2 billion fund to back high inflection investments.
There hasn't been a better time to start a company and the AI wave may be the biggest yet - bigger than the
A decade old throwback to setting foot at the
@kleinerperkins
offices for the first time as one of the 2014
@kp_fellows
.
Still the best program out there for college students, hands down.
Despite outward enthusiasm, businesses want better software guardrails for LLMs before adopting them more broadly. Without more work on building compliance standards around models and applications, enterprise adoption will be slower than expected.
Published some thoughts on the
Someone should reboot Pocket, where saved articles are auto-summarized, there's an option to listen to articles and converse with them while walking etc.
The general anatomy of the new crop of cloud databases:
1. Serverless.
2. Written in Rust.
3. Separation of storage and compute.
4. Support of incrementally computed views.
Two fascinating paradoxes in enterprise software today: UiPath & Splunk.
Both have solid financials and continue to grow well. Yet they both:
1. aren't loved by customers
2. are quite clearly facing an innovator's dilemma.
It's nice to see Palantir's FDE model become popular amongst the current crop of AI startups.
In fact, it's becoming a critical part of the sales motion given that buyers not only have slightly custom needs from LLMs, but also want vendors to just make things work with some
Looking back, Palantir played the recruiting game masterfully.
Lots of things that came together but I’ll call out a time when referral bonuses for technical talent were $10k-$25k that could be accepted either in cash or equity.
Takeaway for me is that growth stage cos with
Databricks' announcement today fully reveals their aspirations to compete with the cloud providers and OAI for the next wave of LLM driven compute -
Lakehouse AI is single handedly taking aim at the LLM Ops stack, HuggingFace and even OAI (curated models
My Twitter feed took a breather from AI to inform me multiple times that the captain of the US cricket team is an Indian immigrant who is also a principal software engineer at Oracle.
So ready for this summer of cricket 🏏
Last week we brought 150
@kp_fellows
, across 10 classes, together in NYC for a fun night of conversation and catching up.
Was personally great to reconnect with other '14 fellows and hear from two amazing founders,
@bradmenezes
and
@narayanarjun
, about their inspiring journeys
We need to move past just vector embeddings to make progress with LLM reasoning. This article on GraphRAG is what I've been looking for:
Not sure where the final architecture ends up - do concept graphs complement vector stores or does encoding move
While data privacy concerns have always been in the ether, they appear to be materializing more vehemently under the backdrop of AI.
Growing number of cases where startups can’t close their initial set of customers without supporting self hosting (on-prem / VPC) out of the
The promise of LLMs applied to dev productivity is not in code gen, but in supercharging the tedious:
- Codebase navigation + understanding
- Test writing, refactoring, tech debt elimination
- Documentation generation and summarization
- Debugging
Devs love their core jobs!
Cool to learn that stable diffusion was motivated by an engineering problem - build a computationally efficient diffusion model.
This led to the core insight of decoupling the training of the auto-encoder model and diffusion model - which is arguably pushing the science forward
This Snowflake hack goes to show that for all the complexity in the cyber stack, simple things still go unnoticed.
All the attacker did was phish login credentials and then use a stale refresh token from the inside to generate various session tokens that allowed them to exflil
Feels like Nvidia has an underrated opportunity to build a monster cloud biz.
They’ve been building more and more software primitives over the years tightly coupled to the hardware. Why not double down?
Apple-esque chance to build a walled garden around next gen GPUs and
Had fun prototyping a
@dust4ai
app that answers a couple of popular software engineer interview questions with code as output.
h/t
@spolu
for adding slick integrations &
@KunhaoZ
whose template I tweaked.
Such talk feels overblown. Why?
1. Inference costs need to reduce 10x. chatGPT bleeds ~$0.05 per call. Google gets 8B queries a day.
2. chatGPT doesn't surface real time info. Retraining at scale is necessary and costly.
3. Needs biz model innovation.
4. Google's brand moat.
"Paul Buchheit, the former Google employee who created Gmail, wrote in a series of posts on Twitter that (post ChatGPT) the company may be “only a year or two away from total disruption.” $GOOGL
Can’t believe it’s been 11 cohorts of
@kp_fellows
! Had one of the best summers of my life in 2014 as a fellow and encourage anyone who’s interested to apply.
We also get together and do fun things. If you wanna meet some of us and hang out, just DM me.
’s such an underrated SaaS company. Remember seeing ads for it plastered in NYC subways in 2015. Didn’t have any friends who used it or were aware of it.
It now generates over $1B of ARR, growing 35% YoY. 88%+ GM and ~30% FCF margins in 2023.
After two quarters at business school, I have a greater appreciation for the variety of "smarts" that people in my class possess. Book smarts, technical smarts, smarts about navigating systems, smarts about navigating conversations and relationships, domain smarts. All valuable.
Just watched a Stanford student pitch his startup. The words "AR", "blockchain", "tokens", "fractional ownership", "commoditization" and "disruption" were uttered in the first 2 minutes. Watch this company raise a 30 million A round.
So pumped!
@jarredsumner
is special and it’s no secret that Bun has taken the developer community by storm. Best part is that they’re only just getting started towards an extremely bold vision. Needless to say, they’re hiring!
We
@kleinerperkins
are excited to announce our partnership with
@jarredsumner
and
@oven_sh
as part of their $7m seed.
Oven is building serverless hosting and CI for JS apps, powered by .
(1/2)
These are the financials of….
A non-venture backed cloud kitchen called “Mylapore Express” that serves homemade Indian vegetarian meals to just Bay Area residents.
We have
@anaganath
from
@kleinerperkins
on the Infinite ML pod today. We talk about AutoGPT along with a range of topics including:
- ChatGPT Plugins
- Context windows of LLMs
- LLMs for developers
- AI landscape and investment opportunities
After playing around with selenium for a side project, it's fascinating to think about just how many RPA use cases can be solved with some LLM prompting -> real time code-gen wielding a headless browser via selenium.
Of course, bunch of tweaks required to sell to enterprises +
A cool tidbit on Figma was that
@evanwallace
was even optimizing chrome source code (example: how graphics values are converted) to ensure figma was performant in the browser.
A lot of low level, systems heavy work was equally responsible for Figma's product success.
Consumer AI product I'd happily pay for:
System that calls customer service on my behalf and negotiates reservations, cancellations and reimbursements. It needs to:
1. Have key personal details -> handled in the onboarding flow.
2. Ingest situational context -> provided by me.
@HarryStebbings
chat with
@lennysan
is amazing with many great nuggets. Here’s one I found hilarious:
Harry: What are signs that a PM from Google or FB would be able to work well at a startup?
Lenny: They haven’t been there too long or they’ve worked at other companies too.
Thrilled to announce I’m (re)joining Kleiner Perkins as a Partner to help lead our inflection/growth-stage investing strategy. I have an immense amount of respect for this team’s approach to & track record in backing history-making founders, and I’m stoked to be back on team KP!
Two thoughts on AutoGPT as it gets more popular:
1. Would be great to be able to intervene "mid flight" and either help it out when it's stuck or change direction based on dynamic preferences.
Otherwise you're limited to tasks where the constraints are static and can be
Will be interesting to see if Jasper continues down its path of building its own custom models vs. switching to GPT-4.
In many ways, Jasper is a useful leading indicator to understand behaviors and implications to Open AI's business model.
Bought an iPad pro along with the Pencil, and it might just be the single most important purchase I've made in years for my productivity. Notability on the iPad is amazing and importing all of my readings and lecture slides into it so that I can annotate on top of them is insane.
Open-core is a fascinating business model because it gets *harder* to monetize the *better* your open-source product is.
You can create inorganic distribution for a mediocre (high time to value) OS product if you're first to market & have strong top-down GTM.
An under appreciated discussion point is how incumbents did “free” market education / category creation for AI startups in certain categories:
1. GitHub copilot showed the value of a coding copilot to enterprises in 2022, paving the way for startups to seize the opportunity.
2.
With Canva raising a private round valuing it at $40B, it’s worth noting that Australia has produced some amazing, global companies:
Canva: $40B MC
Atlassian: $100B MC
Afterpay: Acq by Square for $30B
Definitely an underrated "startup" geography relative to India and Israel.
Interesting tidbit from a conversation with a founder:
Feeding Sora 360 videos to a Gaussian Reconstruction Model to recover 3D assets.
Makes me wonder if Sora should be thought of as a platform vs. as a model.
@darian314
Yes.
@KurtosisTech
,
@_hex_tech
, Arena AI,
@partifulsocial
are some examples that come to mind. Many more emerging.
Palantir hired smart people and exposed them (including SWEs) to users repeatedly, in a high pressure, fast paced environment.
First time in Manilla and very impressed by how modern everything is - virtually no pollution (including noise), great infrastructure, bustling nightlife and lots of foreign brands.
Next stop, Taipei!
If one believes that each knowledge worker will have their own, personalized and private LLM, it’s interesting to think about how that could be delivered by a 3P vendor:
1. (Obvious) each person has access to an isolated model.
2. (Less Obvious) what does true multi-tenancy
Very sad to see so many posts of people having their job offers rescinded. Especially sad for internationals whose US visas are tied to their employment.
@kleinerperkins
has plenty of portfolio companies that are actively hiring and I’d love to help those who are recruiting.
Wow, Geoff Hinton just won the Nobel Prize in physics for his work on neural nets.
First computer scientist to do so. He is also a Turing Award winner. A historic moment in the world of academia and an even grander acknowledgement of deep learning's fundamental unlock.
The innovation battles within "AI" are interesting and good for the ecosystem. Two prominent ones:
1. Open AI vs. Open Source Model Ecosystem: OAI is forced to innovate (GPT-4, ChatGPT) to maintain distribution advantage.
2. Github's Next vs. AI dev productivity startups
Datadog moving into cloud service management with AI makes so much sense. Lots of depth and value in those workflows.
Meta points:
1. Even a typical infra company like DDOG is thinking hard about its "AI angle".
2. System of intelligence is the next frontier of observability.
@jebank
That’s right. Along with the mystique of what they did, branding of how they were solving hard problems at a time when consumer social was hot, the A+ candidate experience that felt luxurious and the folklore around the founders including Thiel.
Enjoyed
@parkerconrad
's interview on
@InvestLikeBest
Reinforced the following:
offering point solutions bundled together != compound startup.
You need a foundational layer which can either be a data integration play (Rippling) or a compute bedrock (Databricks - Spark).
@semil
@Austen
Austen is super active on Twitter because he realizes that at the end of the day, Lambda's moat is its brand (which he is an extension of). Don't think this strategy is that effective for founders of traditional consumer tech software companies.
Eagerly awaiting a virtual assistant that I can text, which manipulates a headless browser behind the scenes to take complex actions:
1. Booking travel
2. Making reservations under multiple constraints
3. Completing multi step tasks on annoying websites.
@martin_casado
Shouldn't it be the opposite? Hot startups haven't had down rounds while public cos have been slashed.
Same public cos offer fixed $ worth of RSUs, which talent would be happy to scoop at today's prices?
Cool! What I like about this is the design decision to not compete with writers but instead supplement them.
Broader point: "AI first" tools that supplement rather than supplant humans will do well in the coming years.
Introducing Lex!
A word processor with artificial intelligence baked in, so you can write faster.
👉 👈
(I've been working on this awhile... so glad to finally share it!!)
The merits of an open-core approach around a language model is a fascinating open question, as far as I know. A few thoughts on monetization paths below -
The rate of diffusion of this next generation of AI is unlike anything we've seen, but even more remarkable is the sense of empowerment it has already unlocked in every corner of the world, including rural India.
If crypto could emerge out of fintech and be a standalone category that funds staff and raise against, only a matter of time until AI does the same vis a vis enterprise software.
LinkedIn was designed to better facilitate professional connections. However, most people around me use it to find others via search and then use email finding services (Rocket Reach) to form connections outside the platform. Email has a poor hit rate but > LI DM hit rate.
Retool replicated a single microservice in Palantir's platform (15/20 services & data integration + storage capabilities) and is already 1/16th $PLTR's market cap. Seeing a lot of other startups attempting to unbundle Palantir in the enterprise.
@chamath
If Dems win, it's a particularly nice boost to commodities + energy (nuclear etc).
Given today's price action in gold, silver and uranium, the market seems to have a prediction :)
How would we hear this graph of exoplanet WASP-96 b’s atmosphere? The pitches of each data point correspond to frequencies of light, with longer wavelengths having lower pitches. Volume indicates the amount of light detected. Four droplet sounds represent clear water signatures.
ChatGPT is awesome but like GPT-3, it can't do math.
While GPT-3 just outputs a wrong answer, ChatGPT tries to sound authoritative while being wrong.
h/t to my colleague
@haomiaoh
Nice read. Meta point is that we still need to make a leap in reasoning capability for new markets to be unlocked.
Also likely need to have agentic products come with built in steerability, async-await equivalence etc.
Wonder if agentic platforms can perform lookups on
Dune 2 was a spectacle with tremendous cinematography. Feels like that's where movies continue to head.
While deep storytelling / character development has been offloaded to shows.