Kevin retweeted
A slide from @AnthropicAI's IPO roadshow. Hits hard and is very true.
AI is soooo much better than how most people are using it.
Kevin retweeted
AI Data Companies: Thoughts on Org / Comp Structure
I've been thinking about the right org structure and compensation philosophy for AI data companies.
Rough thoughts: (THIS IS NOT LEGAL ADVICE)
1. Avoid incorporating as a C-Corp due to double taxation of dividends. Instead, create a S-Corp or LLC.
2. Don't raise venture money, at least initially. Raise money from family offices or angels since VCs can't invest in a S-Corp or a LLC. If the company does turn out to have durable equity value, you can always switch to a C-Corp (S-Corp might be better for this reason than LLC, but pls consult your lawyer).
3. Raise the amount (+ buffer) you need to get to cash flow positive (I think this is likely sub-5M, probably even 1-2M). Obviously this is easier said than done for most, but if you have a strong background and a differentiated POV + lab connects, there are TONS of folks who'd be interested in this kind of structure.
4. (this will be controversial) Don't give any equity to employees. Everyone should be paid a relatively standard base + a highly variable bonus, which is a % of gross profit they are tied to, similar to a hedge fund.
5. After all employees are paid out, keep 30% of the remaining cash on the balance sheet and dividend out the rest to shareholders.
Re org structure, run the company like a multi-manager hedge fund. (a) small P&L-owning pods per lab or vertical (b) a thin shared platform (c) an independent QA function with veto power, since QA is SUPER important (d) a tiny full-time core, and (e) an elastic expert network of contractors.
In general, I think a hedge fund style org and comp structure is the best way to go here. Most hedge fund employees don't own equity in the underlying management company, but nobody ever complained of not owning equity in Jane Street.
AI data companies are similar to hedge funds because:
(a) The edge decays. A hedge fund’s alpha erodes as others find the same trade. A data company’s edge erodes too: today’s frontier data (a new domain, a new eval format, a new kind of expert annotation) becomes commoditized within a year or two, and labs bring capabilities in-house or models learn to generate it themselves.
(b) Returns come from people, and you can attribute them. At a fund you know which PM or desk generated the P&L. At a data company you know which team owns which lab relationship, which project, which data vertical. When contribution is measurable, a cut of gross profit is a fairer and more motivating incentive than a slice of a cap table whose value nobody can predict.
(c) It doesn’t need much capital. Like a fund’s management company, a data business is mostly people and some tooling. Once it’s cash-flow positive, more capital doesn’t make it grow much faster. Growth is limited by finding great people and earning lab trust, not by money. Raising a big round just dilutes owners to fund spending the business doesn’t need.
tl;dr if the business looks like a high-margin, people-driven, cash-generating machine with uncertain long-term moat, you should structure it like one. Optimize for distributable cash and paying the people who generate it, and keep the option to become a “real” equity story if one emerges.
Kevin retweeted
a few of my predictions for the next 6 months of ai data:
- a monopolistic human data buy side layer emerges. right now you can use zap mail, heyreach, instantly and claude max plan to reach out to practically every potential human data buyer on the world in 7 days and (even with the worst strategy) at least come out with roughly 20 buyers spanning rl builders, neo labs and (if lucky) one or two major labs, who you could easily offload $20-$50m cumulatively from the very start. it is not going to be this easy to find buyers forever. as labs inevitably get f*cked by bad apples on rights, de-id, QC and the rest, a select few will emerge as “transaction gates” in which every corpora must flow them in order for the labs to deem as safe to buy. analogous to how a few labs already don’t touch corpora if it was de-id by Tonic Textual.
- the waterline continues to rise. ai labs are buying real company data to build artificial workspaces to teach their models how to do an office job. using a headcount as a metric, 3 months ago they bought company internals with 10+ headcount, now they almost exclusively purchase from companies with 50 or more employees. as models saturate these smaller workspaces, demand for smaller company corpora plummets whilst demand for the already dwindling supply of large sprawling company internals increases. yet doesn’t skyrocket as at a certain point if the price becomes too high then labs fall back on synthetic expert work. it goes without saying that small companies in a esoteric domain models underperform in that hold vertical depth (e.g a 30-man freight brokerage) will still hold immense value.
- reality gets graded more and more. as the most knowledgeable domain experts begin to fail in writing valid rubrics that aren’t instantly saturated we are going to have to get reality to play the verifier, domain expert and environment, if we want progress to continue. rubrics shift from the attributes of a perfect NDA to a lab measured critical tempurature.
- atom environments. why hasn’t the vertiginous increase of ai capabilities led to huge economic growth? how can we get non-tech people to understand how intelligent ai is already? how do we find the best value/token ratio? to “solve” these questions I believe need to give ai a “body” in which it can interact with the real world at extremely low bandwidth. giving ai a body sounds like a humanoid which is obviously just one small slice of real world interaction, a little bit like how literally none of the ways we interact with in the real world is only via our real body but via tools instead like handles, vehicles, instruments and cutlery. maybe I should reframe it from giving ai a body to giving the world ai
- protocols for company data emerge. for such a fuzzy yet crucial area of ai, it is insane how unorganized certain steps in company data brokerage are, such as giving context on what your companies internals actually hold.
Kevin retweeted
I've been thinking a lot about data bottlenecks, and I'm not sure who else is screaming this from the hilltops. I feel like I'm taking crazy pills looking at the jagged frontier thinking it's smooth if you order by data volume.
the spikes in the jagged frontier are tasks with tons of data (either static or simulated; think math, programming, games), and the valleys are tasks with little or no data (cooking a lasagna, olfaction, manufacturing a passenger train). I don't know of any exceptions.
if you order the jagged frontier by data it looks like a spiral frontier. models are superintelligent where we have enormous data, and progressively less so as we have less data. some tasks may draw data from static lakes of data, some may be from limitless waterfalls from simulators that convert compute into data. but ultimately, intelligence on a task is proportional to the data absorbed by the model for that task.
scaling laws are real. if you are not bound by compute or algorithms (increasingly we are not), then intelligence is proportional to data. since we don't have a very good notion of task similarity, it's often difficult to know if data for one task will help you with another, without training a model and seeing. but, it doesn't seem like chess data helps much with math performance and vice versa. so when I refer to data, I refer to data that is task-relevant.
maybe we call this the 'no shortcuts hypothesis'. all the rules-of-thumb about learning are at the root, the same -- Malcom Gladwell's 10,000 hours, Moravec's Paradox, Scaling Laws, the Bitter Lesson, Age of experience. If your learning system (model, brain, whatever) is unconstrained by compute, intelligence is proportional to accumulated experience. no shortcuts. want to be superhuman at chess? play all the chess, more than the whole species combined (good thing a model can do that in 24 hours). want to be superhuman at math? build a math simulator and lock the model in it and run it faster than reality. superhuman at coding? first, collect all public examples of humans writing and responding to code ever, make the model learn from that first, and then lock the model in a simulator where it fixes bug after bug, faster than realtime.
superhuman at image generation? no shortcuts. but here, we don't have a simulator. so now we're moving around the spiral frontier away from closed-loop into large static lakes. still big enough to be superhuman though, which also applies to video, text, and audio. oh! and once you have a superhuman model, you can use it to bootstrap other tasks by making a superhuman model a judge on other tasks.
let's keep marching along the spiral frontier, out of the realm of data abundance, simulators and large static lakes.
biology? no shortcuts. go get a ton of data. and be freaking careful! you can get bad data in biology pretty easily. tilting a well plate 5 degrees the wrong way can ruin an experiment. so make sure you understand the domain, otherwise you're going to have to filter a lot out.
robotics? no shortcuts. get the data. if you can build a simulator, do it, but be mindful our simulators don't have perfect correspondence with reality. so, build an arm farm. fold 1M shirts, put 1B balls in cups. play ping pong forever.
olfaction? no shortcuts. smell the world. build a laboratory of chemical sensors, an international network of trained sensory panelists, a scent production factory, and link them all together. make millions of samples. smell millions of samples. take billions of chemical sensor readings.
I totally buy super intelligence as a concept because I feel it, I use it for coding, the breakthrough math results speak for themselves. I see RSI coming, because how could an all-digital task like ML research and model building not fall to the current paradigm? But the notion that AI will come for everything, all at once, irrespective of domain, feels silly, because there are no shortcuts. if you want a specific form of intelligence, you have to get the data.
looking at you @willdepue, I feel like we're both seeing the same thing
🤖 Made with AI
Kevin retweeted
As autoresearch agents run longer, do they keep exploring new approaches, or mostly refine the ones they found early?
Turns out, runs settle into different "idea basins" and stay there.
We introduce fork-and-flush, a simple way out 🧵👇
Kevin retweeted
Increasingly becoming convinced the only value left will be in the 5% ai can’t do for a few decades well (at least).
1. Physical.
2. Strategy and very novel ideas.
3. Complex compliance moats.
4. Brands.
5. Community.
6. Authority and trust.
7. Community and events.
Dollars will shift around accordingly.
Kevin retweeted
I work in the AI industry. I burn about 1.5 billion tokens a week. Between personal and business I have 8 AI subscriptions. And I still think I am falling behind the accelerating rate of change. The last three months alone have been a quantum jump in capabilities. Nobody outside the industry understands how insane this is going to be. Like a whole other industrial revolution, but compressed into 10 years. Nobody is ready for the societal implications. It's going to be a 10.0 earthquake plus an F5 hurricane all at once, globally.
Kevin retweeted
Nobody actually needs a personal AI assistant. The average screen time in the US is 7 hours/day. people have plenty of time. There's no evidence that freeing up more will lead to anything but more tiktok scrolling
If anything daily life needs more offline friction
Kevin retweeted
If you raw dogged a CS degree pre-LLM you can never be a true vibe coder. You’ve seen too much.
- 90% of work in a company is classification, “best next step”, reviewing
- video game AI is about to step change improve, accelerating simulated worlds for RL faster
- chaining Jev into a text model for final answer might be the move
- is Routing over or does it need to evolve?
Some raw / unfiltered thoughts around the Jev model. Thinking out loud:
- Lots and lots of problems in business are classification or regression problems. Lots and lots of companies don't know that the types of problems they have are solvable by classical ML methods. They often solve them with people and process instead of technology.
- With the emergence and popularity of LLMs, more companies are thinking 'Maybe we can use AI for that!' and are solving classification and regression problems with LLMs.
- This is good in some ways, because companies are potentially automating some manual work, but also bad in some ways, because it's often the wrong tool for the job, and possibly not as good as the classical ML methods for what they are trying to do..
- If anything, the emerging popularity of 'AI' (which for most folks is synonymous w/ LLMs at this point) has caused them to try to solve their problems with that type of 'AI' (even if classical methods are more appropriate if you knew about the available tools / techniques)
- There are very established 'classical ML' patterns to solve the problems they have (e.g. XGBoost is often all you need). AutoML works pretty well at this point for the 80% of problems here too, as long as your data is clean.
- But the classical techniques require you to label your data, train a model, and host that model somewhere. They aren't as easy to use compared to 'calling an LLM API' - and it requires you and your org to be aware of those techniques, and capable of investing in them.
- Classical ML is now much easier to do because of LLMs. I've built a bunch of models in the last 2 years using LLMs to build the classical ML code, do data cleaning, etc. - that would have taken me *much longer* if I didn't have an LLM to help - but still harder than shoving the problem at a text-based model.
- There are some great patterns / techniques which exist around LLMS to constrain output for classification / regression, and make things cheaper / more performant (Structured Outputs, DSPy, GEPA, etc.) - still not super-accessible to the broader audience (I can count on one hand the number of practioners I know who have experimented with DSPy / GEPA -- maybe a few more who use Structured Outputs) - I am hopeful that these types of things become more adopted, but maybe the technical hurdles are too high in the short-term.
- With Jev, maybe they've nailed the 'UX' for this. Yes, it looks a lot like structured outputs, but it is more opionated and directly targets classification / regression -- in a way that probably more folks will understand vs. Structured Outputs / DSPy / GEPA
- I do wonder how effective it will be. It feels very adjacent to zero-shot Timeseries stuff like TimesFM. Like yes, it is very cool you can do predictions without training, but when you look at stuff like TimesFM, it seems to be not very good compared to supervised training of TS models (again, requires expertise to do so). Maybe you can put your labeled data in context, and this will work
- But then my question becomes -- if it takes a bunch of know-how to use this type of zero-shot classifier / regressor model, at what point is it better to just use LLMs to help you build classical ML models?
Thank you for reading my rant. I know it is disorganized and probably not fully coherent. Either way, it's super-cool to see new ways to piece together this new tech that has emerged over the last few years. Excited to track where Jev goes, and I think it prob. will be useful for some. I plan on trying it out for sure.
1
3
90
Kevin retweeted
After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI?
I’ve spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev
• 20-200x faster
• 40-400x cheaper (w/ output tokens free)
• Frontier composable intelligence optimized for decisions
AFAICT the shortest path to AI-based economic revolution
Kevin retweeted
bullish on data companies going into training, it’s so obvious that data vendors need to understand their data and learn if it is actually good through training
highly doubt that any of these "data labs" will post train anything useful or interesting but I expect that there will be some interesting research that come out in the next few months around data
even internally, post training teams spend a good chunk of time trying to understand what data to train on and what capabilities the model is learning, this is work that companies supplying data should know better than us
[note: any data vendor that outsources training to someone else is retarded, most of the value is in the process of training itself. secondly, training on tasks that are difficult for GLM/Kimi is quite different from what is happening at the frontier; if you think Ant/OAI don't have alpha relative to open source in how they train Mythos/Astra you are wrong]
companies im bullish on to do this:
- Preference Model
- Fleet
- Proximal
- Mercor*
*Mercor is making a big push here, I expect some interesting things to come out from them, but still skeptical since most of their work has been slop
companies im bearish on to do this:
- Afterquery
- Handshake
- Datacurve
- Surge
- Every other YC rl env co
Instinct, the viral AI assistant, has already raised $350 million, including a new round that values the startup at $2.5 billion. wsj.com/tech/ai/the-latest-v…
1
249


























