
Building @try_whitney | alaskan | prev @cloudcruise_app (yc w24), @microsoft
San Francisco
Joined October 2024
- Tweets1.2K
- Following878
- Followers560
- Likes5.4K
World Models are slated to be the next frontier after LLMs.
@vatsalbajaj and I had the pleasure of chatting with @sethia_avik , Founding Engineer at @reactorworld , who recently raised a $59M Seed and Series A to be the Developer Platform for World Models.
Tune in to find out what world models are, the unique challenges behind serving them compared to LLMs and exciting future applications!
Chapters
00:00 - Introduction
01:09 - What is Reactor?
01:57 - What is a world model?
05:10 - Interactivity differences from Fei Fei Li and Yann LeCun's world models
12:47 - How world models are used today
15:03 - Challenges of serving world models
19:10 - Differences between serving World Models and LLMs
21:17 - KV Cache Challenges
28:06 - How LLM serving experience translate to world models
30:00 - Architecture differences
34:56 - Open Source World Models
39:35 - Robotics Models
40:51 - Future Applications
Jake retweeted
Yaroslav Bulatov (@yaroslavvb) wants to get rid of backpropagation.
For Episode 7 of The Interlude Show, @jakekang and I chat with Yaroslav, who worked at Google Brain, early OpenAI, Meta, and Together AI. He now runs the Sutro Group at South Park Commons (@spc), exploring energy-efficient learning algorithms with AI agents.
We get into why moving data matters so much for energy use, how the memory wall changes algorithm design, and how gradient checkpointing can make a workload faster even while doing more calculations. We also discuss early OpenAI, his skepticism about reinforcement learning, and why recursive self-improvement still faces new bottlenecks.
We close with whether more code translates into better outcomes, and the difference between AI that helps us pursue our goals and AI that gets better at capturing our attention.
Chapters:
00:00 - A plan to rethink backpropagation
00:49 - Introducing Yaroslav Bulatov
01:30 - Energy-efficient learning and the 20-year research plan
03:25 - Why focus on software now?
08:05 - Sutro Group: using agents to explore learning algorithms
10:57 - The memory wall: arithmetic versus data movement
17:07 - Why peak memory is only part of the energy story
18:40 - Why Yaroslav wants to replace backpropagation
24:27 - Mixture of experts and disentangling neural networks
27:24 - Toy problems: matrix multiplication, sparse parity, and MNIST
31:21 - Early OpenAI and skepticism about reinforcement learning
34:41 - Gradient checkpointing: doing more math to move less data
38:51 - Power grids, intelligence per joule, and new bottlenecks
43:25 - Do more tokens and pull requests mean better outcomes?
46:21 - Superintelligence, specialization, and physical limits
51:49 - Creative freedom, attention capture, and possible AI futures
Jake retweeted
Traces are the new oil... or so says my X feed.
But how are people using them? Why are they valuable? How are they connected to training, or evals?
@vatsalbajaj and I are hosting @ankush_gola11, co-founder and CTO of @LangChain, for a fireside next Thursday to explore questions like these. Through LangSmith, LangChain sees more production agent traces than almost anyone. Now they're moving to help companies turn those traces directly into better models.
Perhaps there is more to life than buying and selling RL environments. Perhaps your own traces are worth something, too. :) Come learn with us: 10/22, in-person at @spc, 3-4pm. Doors open at 2:45pm.
luma.com/posttraining-langch…
Jake retweeted
The frontier data and post-training industries are starting to converge.
I recently hosted @BrendanFoody at the @spc post-training forum for a fireside. @mercor went from $1M to $2B+ run rate in ~24 months selling the data labs train on, so I wanted to know where he thinks data fits in the post-training story.
A running theme at our forum is that post-training is nearly all data: building evals that capture what you want, and wrangling data to hill-climb them. Past conversations have put that at "99%" of the work. And it's hard work.
But it turns out that convergence runs the other way too. As labs need harder RL tasks, a frontier data company can't just hand over expert annotations. It has to prove the data moves a model. Post-training becomes a vital component of the data and environment curation process itself.
"For us to know that a dataset is useful, ultimately we need to be able to train on it." So Mercor hired Edward Hu, first author on LoRA, to lead post-training. The best data companies are getting pulled into post-training, just as post-training companies have found themselves drilling deeper into data discipline.
I think this is a telling glimpse into the future of AI in enterprise. Enterprises today have proprietary data, but proprietary data alone doesn't get you far. A model can't learn from reading your Slack any more than you could. The data has to become evals, verifiers, environments - before you even get into the further decisions of base models, algorithms, and harnesses. Building or partnering with talent that can turn data into training signal will become essential.
Early software created an explosion in demand for programmers. We're at the start of the same boom for post-training.
Find the full transcript of the conversation on the @try_whitney site: trywhitney.com/journal/data-…
Jake retweeted
"The entire way that humans contribute to the economy is going to move.”
@mercor Co-founder & CEO @BrendanFoody explains what he believes will replace repetitive human work.
I've talked to so many RL env founders in the past couple weeks, many of which have really interesting data / traces but are unsure how to best present it for labs.
playbook seems to be:
1. build a benchmark
2. baseline models
3. post-train a model
4. inbound from labs
5. $?
1
4
306
Very excited for this talk - data is 99% of the work in post-training. You need high quality data and you need data that'll improve your model in areas you care above.
Will the frontier labs relaunch fine-tuning products? What differentiates data suitable for post-training and data that is not? What does recursive self-improvement mean for the value of human expertise?
I'm excited to dig into these questions and more in a couple weeks with @BrendanFoody , CEO of @mercor, the company behind the data and RL environments powering most frontier labs. This is the latest speaker series of @spc's post-training forum: in-person, live, 9/16 from 3-4pm.
Data is often the most central - yet least understood - component of model training. Arguably it matters more now than any architectural advance. Come find out why.
luma.com/spcforummercor9-16
4
280
Jake retweeted
Humans can dissect their own thinking instinctively, but models need expensive post-hoc interpretability. Chain of thought may not be a faithful record of what a model actually computed.
Jasper Gilley (@0xjasper) researches brain-inspired AI architectures at @spc, particularly what the cerebellum suggests about giving a model access to its own computation. Previously, he worked at Yutori where he built vision-based web agents that emulate human users.
Check out Episode 4 of The Interlude Show where @jakekang and I chat with Jasper. We get into forward self-models (predicting a transformer's own activations), Zipfian Grokking, why Jevons' Paradox won't save your software engineering job and what humans were evolutionarily designed for.
Chapters:
0:00 - Cold open
1:02 - Intro
1:43 - Researching brain-inspired architectures at South Park Commons
4:05 - The cerebellum as a forward model
6:39 - The Forward Self-Modeling paper
14:03 - Grokking, explained
18:37 - Zipfian Grokking: finding the solution and forgetting it
22:26 - Yutori: why vision beats the DOM
31:23 - Why Jevons' Paradox won't save your software engineering job
41:51 - The window where creativity is the whole job
45:39 - What humans were evolutionarily designed for
51:09 - The future of post-training
Post-training in cybersecurity is an emerging field.
@CrowdStrike published a paper last month on training a SOC triage classifier where optimizing on accuracy alone gave them a model that scored well but reasoned poorly - rewarding only the final answer can often push a model toward the wrong reasoning path.
Reward attribution / design is one of the hardest parts of post-training. Check out the post below on how we thought about turn-by-turn attribution over scoring entire trajectories, and how a 27B model beat off-the-shelf frontier attackers on success rate at lower cost and latency.
Always happy to chat with anyone working on post-training, cybersecurity, or both!
Last week, @OpenAI published an open letter, signed by over 200 companies, calling on security teams to continuously test their defenses against frontier attacks.
Threats like prompt injection have evolved from single-turn injections to sophisticated multi-turn campaigns.
To defend against these attacks, teams rely on red-teaming, simulating real-world adversaries to test defense readiness. We partnered with NuGuard AI to post-train a red-teaming agent that outperforms frontier models at lower cost and latency.
We post-trained Qwen3.6-27B using a modified GRPO loop that attributes reward turn by turn instead of scoring the whole conversation at once. The specialized model lands 3.1x more attacks than frontier alternatives like Sonnet 5 and DeepSeek-V4-Flash.
The attacks in the write-up are illustrative, published to advance defensive research. Full methodology in the comments.
2
7
515
Very fun conversation chatting with Ryan, who's very knowledgeable in the space of healthcare!
Jake (@jakekang ) and I interviewed Ryan Tsang (@rvtsang), who is building Harmonize Health - an AI-native operating system for medical clinics.
Ryan spent four years at Microsoft working on Azure OpenAI's fine-tuning product, then trained models for open-source AI safety benchmarks at MLCommons.
We get deep into how Harmonize Health builds and grades evals for voice agents (including a user-simulator agent that role-plays the patient calling in), why you should get to production before you build the perfect eval set, how Harmonize scrubs patient data out of transcripts, and the tension between forward-deployed engineering and scaling across multiple clinics.
Hope you enjoy the third episode of The Interlude Show!
Chapters:
00:00 - Move fast and break things doesn't work in healthcare
00:41 - Meet Ryan Tsang and what Harmonize is building
02:40 - Why healthcare, and the path from pre-med to CS
05:45 - Inside a primary care clinic: the work nobody sees
10:50 - Safety is the hardest technical problem
13:55 - Data plumbing vs. AI: EHRs and tribal knowledge
20:36 - Encoding a clinic's operating procedures into agents
23:35 - Inside their eval harness for voice agents
27:07 - What can we really not screw up?
30:16 - Healthcare moves at the speed of trust
38:57 - TempShot: pitching his high school app to us as VCs
49:53 - The YC all-nighter, quitting our jobs, and why compounding wins
Jake retweeted
Replying to @KJHMiao
@KJHMiao and I held a post-training fireside chat with the @trajectorylabs team to discuss their vision for continual learning.
I particularly liked the distinction between "experience" and "IQ". Another strong reason why so many firms these days are emphasizing the importance of AI that you own!
00:00 Intro
00:34 Where the continual learning vision came from
07:03 Why a platform instead of forward-deployed engineers
14:23 Where the name Trajectory came from
19:53 Labs optimize for IQ, we optimize for experience
25:14 Continual learning without touching the model
32:06 Hot take: the most underrated part of post-training
36:45 AI natives vs tech natives vs enterprises
42:58 The magic moment: wake up and it's smarter
45:09 Three possible worlds
Jake retweeted
I built ML Serving and Inference architectures at Netflix and Microsoft for 4-5 years.
A friend of mine (MLE @ Doordash) pointed out how far behind the US is on generative recommendation research as against China. As I followed the trail of research papers, I am more convinced that we need to focus on not just modeling but also applications of agents and LLMs. AI adoption is the biggest risk today.
In this episode, @jakekang and I get into personalization, generative recommendations architecture, inference / serving optimizations and China vs. US on ML research. Hope you enjoy the second episode of the Interlude Podcast!
Chapters:
00:00 - Intro
02:18 - How Netflix Personalizes Everything
04:23 - Model Serving vs. Inference
06:19 - Generative & Conversational Recommendations
08:13 - Measuring Success: Metrics & A/B Testing
11:29 - China vs. the US: Open Source & the Road Ahead
17:00 - Risks, Incentives & Regulation
20:17 - Serving LLMs vs. Classical ML & Observability
What does working on agents look like in production?
Started a podcast w/ @vatsalbajaj to talk about our experiences working on production ai systems. In the first episode, we talk about my experience building browser agents.
We will bring on engineers/ctos to talk about their challenges of building AI in their industry.
00:00 - Intro
00:52 - 1st Hire at a YC Startup
02:41 - Building a self-recovering browser agent
04:18 - What are evaluations?
05:01 - 90% of the work is evals
09:46 - Offline vs online evals
12:06 - Harness engineering vs fine-tuning
15:18 - Owning your AI in regulated industries
17:35 - The OpenAI/Anthropic threat
20:52 - The future of browser automation
Jake retweeted
Replying to @nickmecklenburg
@nickmecklenburg and I chatted with @oneill_c , Co-Head of Model Training and @part_harry_ , Research Scientist from @baseten at @spc about the future of post-training.
To lift the ceiling on context windows, they are researching KV cache compaction, and their unique take on continual learning (that it is a context, not a retraining problem) demonstrates their vision on the future of models.
Learn more about Charles' and Harry's views on where post-training and continual learning are headed.
Jake retweeted
"The main thing standing between us and models that learn like employees isn't a smarter algorithm. It's the context window." 🪟
@vatsalbajaj and I had a great conversation with @oneill_c and @part_harry_ on continual learning, the economics of training, and the future of RL.






