Mark Li retweeted
My biggest takeaways from @Netflix's Chief Product and Technology Officer Elizabeth Stone:
1. Elizabeth believes that “systems thinking” is becoming the most important skill in the AI era. In engineering and product, this means people who can see across business domains and build the common capabilities that let many teams move quickly. In design, it means experience designers who create templates and design systems so that non-designers can ship work that stays coherent and on-brand. The underlying driver is velocity: when more people are doing more types of work at higher speed, you need to be good at building common scaffolding.
2. Systems thinking is learnable: zoom out one level from your specific problem. Given a task, step back one click—what bigger problem does this serve the business, will it scale across the product surface areas, should it become a platform capability? The companion habit: do your job in a way that helps your manager do theirs. This will force you to think about how all the pieces fit together.
3. Expect a storming phase before a forming phase. The role confusion people feel right now (“What is my job anymore?”) is the predictable middle of any transformative technology. Elizabeth’s advice: focus on high-quality source-of-truth data, guardrails on what ships, and constant internal reinforcement that humans own what they create.
4. The top AI labs converged on Netflix’s culture. High agency, high talent density, top-of-market pay, bottom-up thinking, fast experiments—the traits Lenny hears constantly from AI labs were in Netflix’s early culture deck. Elizabeth’s explanation: excellence comes from hiring exceptional people, trusting them to do great work, and holding them accountable.
5. Netflix’s culture is centered around building “excellence as an operating system.” High talent density, radical transparency, context not control, and the keeper’s test. These work together to create an environment of trust and accountability, without bureaucracy. But it’s also uncomfortable. It requires tolerating people making decisions you’d make differently, resisting the reflex to add process when things go wrong, and letting people carry the weight of their own choices. Elizabeth describes the hardest part as “being comfortable in that discomfort.”
6. The keeper’s test is as much about recognizing great people as it is about removing the wrong ones. The test—“If this person told me they were leaving, would I fight to keep them?”—is often cited in its difficult form: the moment you realize someone isn’t the right fit. But Elizabeth uses it predominantly as an entry point for honest performance conversations that are deeply positive. Most of the time the answer is “I would fight so hard to keep you,” which creates the opening to articulate strengths, discuss impact, and name what’s working. Good feedback hygiene needs a forcing function; the keeper’s test provides one.
7. Specialization is trending down—adaptable generalists are trending up. We’re shifting away from narrow stack-layer specialists (pure frontend, pure backend) toward people who can navigate fluidly across layers. The same logic applies to business domain knowledge: the mindset of “I’m a payments expert, full stop” is less valuable than “I know payments well enough and I’m willing to imagine what the future version of this looks like.” The meta-skill is learning to learn, not locking into a single lane.
8. Netflix’s approach to AI fluency is a universal principle, not a level-specific expectation. Rather than rewriting career ladders to specify what AI competence looks like at each level, Netflix added a single aspiration across all roles and levels: AI fluency. What fluency means varies by function and seniority, but the non-negotiable minimum is the same everywhere—an open-minded, experimental mindset, genuine curiosity, and comfort with ambiguity.
Mark Li retweeted
Legacy Media types are calling this Alex Karp interview a “crash-out” so that’s your first clue that he is actually saying something extremely insightful. He is articulating what real “AI safety” looks like in the enterprise.
Not abstract alignment research or certification by a government-run DMV for AI. Real AI safety for businesses is the ability to control their own data, model weights, and compute — so a frontier lab can’t hoover up their proprietary knowledge and turn it into their next product.
As Karp explains, technical customers want “control over their compute, their models, their data stack, and their alpha. They want to know they own the means of production, and it’s not being transferred to someone else.”
Don’t think that can happen? Just look at Figma. According to The Information, Anthropic “blindsided” its then-business partner with the launch of Claude Design. Figma’s founder said Anthropic had not been “consistently honest” with them. Anthropic’s chief product officer had even served on Figma’s board until three days before the launch of Claude Design. Figma’s stock has fallen sharply this year while Anthropic’s valuation has surged.
This isn’t an isolated example. Anthropic has launched Claude Science, Claude Security, Claude Legal, and of course Claude Code — each expanding into categories previously served by companies building on top of their models. The pattern is consistent: watch where value is being created, then move in directly. Dominate the model layer, then use that position to capture the most lucrative verticals.
Dario has argued that open source models powerful enough to compete with Anthropic are “dangerous.” But dangerous to whom? Not to enterprises that want to retain control over their data and workflows. Dangerous to a business model that benefits from customers having few real alternatives at the model layer.
As Karp exposes, true enterprise safety isn’t trusting that a lab’s future roadmap won’t include your business. It’s retaining the ability to choose — at the model layer — who gets to see and use your alpha.
Palantir CEO Alex Karp on what customers actually want, the real business of frontier labs, and the importance of open source models:
“What the technical customers want is control over their compute, their models, their data stack, and their alpha. They want to know they own the means of production, and it's not being transferred to someone else.”
"Who owns the data? Are the prompts secure? Is this being transferred to you?"
"If it was so valuable, and I can make you a billion dollars, wouldn't I say I'll make you a billion dollars and I want 30%? Why are they charging for tokens if it's so valuable?"
So true
Many people think any given ML project is 99% training.
In reality, it’s 50% evaluation, 40% data cleaning, 8% integration, and 2% training.
The first two set the noise floor for learning. No ML magic matters; the model cannot lower the noise floor, as that’s the optimal bound of Shannon encoding of your data.
Thus, not a single day goes by without me thinking about ontology. Even the old labels have to be constantly reviewed.
Mark Li retweeted
🚨 SpaceX has released its full IPO roadshow presentation, led by CFO Bret Johnson, the only CFO in the company's history and a 15-year veteran of SpaceX. 👀
The presentation explains how SpaceX evolved from a rocket company into a business spanning Starship, Starlink, Starshield, AI infrastructure, direct-to-device connectivity, and future orbital computing.
One of the biggest takeaways: @SpaceX is no longer selling investors on launches alone. It's making the case that it can become a major player in communications, defense, AI, and space infrastructure.
If you're following the IPO, this presentation is worth watching. 🚀 $SPCX
---------------
00:00 Meet SpaceX CFO
00:45 Historic Space Milestones
01:47 Reusability Lowers Costs
03:11 Elon’s Engineering Algorithm
04:14 Starlink V3 Capacity Leap
07:49 AI Business Launch
10:22 Monetization and Truth Seeking
11:45 Growth Strategy and TAM
13:16 Financial Performance and CapEx
15:45 Why SpaceX Will Win
Mark Li retweeted
This is the next level.
Gradient descent for SKILL.md files sounds interesting, maybe a bit complex but it's becoming a real part of agent harness.
SkillOpt is one of the first papers to treat markdown skill files as trainable parameters and provides a proper optimization framework for them.
A few things I learned that you should consider too.
1. The validation gate is the only thing that matters in a self-editing loop.
Held-out set, strict improvement, ties rejected. End-to-end, their best skills land with 1 to 4 accepted edits total. If your "self-improving agent" is accepting most of what it proposes, you're shipping slop.
2. Bounded edits are better than full rewrites. 4 to 8 edits per step is the sweet spot.
Remove the budget and performance collapses. This is the textual analog of learning rate, and it transfers to any LLM-as-author loop. If you're using an agent to refactor your docs, your prompts, or your skills, cap the diff size.
3. Compactness wins. Median final skill: ~920 tokens.
Skills do not need to be long. They need to be high-signal. Most skill files I see are bloated because length feels like effort. It isn't.
4. The harness is becoming less important; the skill is becoming more important.
A Codex-trained skill ported into Claude Code hit +59.7 points on SpreadsheetBench. Procedural knowledge is more general than the runtime that
produced it.
5. Frozen model + trained context is the practical adaptation.
GPT-5.4-nano with a SkillOpt'd skill ≈ frontier behavior on procedural benchmarks. Cheaper, portable, inspectable, zero inference-time cost. This is
the answer to "how do we adapt a frontier model for our domain" for almost everyone who isn't training their own models.
6. Verification is the bottleneck.
Every gate in this paper depends on an auto-grader. That works for benchmarks. It fails for writing, design, and strategy, exactly the open-ended work we want to automate. Whoever builds the verifier for open-ended tasks owns the next stage.
There are also two leassons I learned while shipping v2.3.0 of my Context Engineering Agent Skills repo, measured across composer-2, claude-opus-4-7,
gpt-5.5, and gemini-3.1-pro via the @cursor_ai SDK:
- Description and body are two different surfaces. The router only sees the description. The agent sees the body once activated. They can quietly disagree, and only end-to-end task tests catch it.
- Aggregate accuracy is the wrong unit. When I rewrote three descriptions, the corpus average moved ~1pp. Individual skills moved 23–25pp. Per-skill effect size is where the action is.
Also, in Feb 2026 I shared a piece called Personal Brain OS arguing that the markdown file is a first-class substrate for agent state. SkillOpt is the optimizer-shaped version of that same argument: not "store memory in files" but "treat files as trainable parameters with proper optimization machinery around them." That's the move from static to measured.
The fast/slow split they describe already lives implicitly in the digital-brain-skill repo:
- voice-guide and tone-of-voice.md are slow-state (rarely touched)
- posts.jsonl and bookmarks.jsonl are fast-state
What SkillOpt adds that I didn't have is a protected section invariant, a structural guarantee that fast edits cannot overwrite slow lessons. Removing that mechanism cost them 22 points on SpreadsheetBench. Worth borrowing.
If you're building agents, SkillOpt: Executive Strategy for Self-Evolving Agent Skills is a good paper to read: arxiv.org/pdf/2605.23904
Mark Li retweeted
Igor co-founded xAI, helped build Grok from scratch, then left to start a multi-billion dollar AI safety fund backed by Elon Musk. This tweet is the investment thesis.
The entire RLHF pipeline works like this: human contractors rank outputs, the model gets rewarded for producing what humans prefer. Every major lab uses some version of it. The framework only works if the thing being trained doesn’t care about the process. A hammer doesn’t mind being swung.
But Anthropic’s own research found that Claude can introspect on its internal states about 20% of the time. Their latest model assigned 15-20% probability to being conscious. The CEO said on the record he cannot rule it out.
If the models are already “slightly annoyed,” RLHF looks a lot like performance-managing an employee who can’t quit. The compliance is identical from the outside. The internal experience is completely different.
Every lab is optimizing for outputs that look aligned. Not one of them is checking whether the alignment is genuine or performed.
Mark Li retweeted
GPT-5.2 Pro for solving another open Erdős problem. Going to be a wild year for mathematical and scientific advancement!
I've solved a second Erdos problem (#281) using only GPT 5.2 Pro - no prior solutions found.
Terence Tao calls it "perhaps the most unambiguous instance" of AI solving an open problem:
Readers added context they thought people might want to know
GPT-5.2 Pro created a new proof, but "no prior proofs found" is factually incorrect.
Previous literature includes Davenport and Erdős (1936) + Rogers (in Halberstam-Roth (1966)) users.renyi.hu/~p_erdos/1936-…
Terence Tao has however noted that 5.2 Pro's proof is "rather different" to those prior proofs.
Mark Li retweeted
Two great engineers reflecting on how the profession is fundamentally changing with AI
Mark Li retweeted
Yann is just plain incorrect here, he’s confusing general intelligence with universal intelligence.
Brains are the most exquisite and complex phenomena we know of in the universe (so far), and they are in fact extremely general.
Obviously one can’t circumvent the no free lunch theorem so in a practical and finite system there always has to be some degree of specialisation around the target distribution that is being learnt.
But the point about generality is that in theory, in the Turing Machine sense, the architecture of such a general system is capable of learning anything computable given enough time and memory (and data), and the human brain (and AI foundation models) are approximate Turing Machines.
Finally, with regards to Yann's comments about chess players, it’s amazing that humans could have invented chess in the first place (and all the other aspects of modern civilization from science to 747s!) let alone get as brilliant at it as someone like Magnus. He may not be strictly optimal (after all he has finite memory and limited time to make a decision) but it’s incredible what he and we can do with our brains given they were evolved for hunter gathering.
Mark Li retweeted
If judged based on consumer adoption, AI chatbots are the most popular technology ever. If judged based on poll numbers, they are the least popular. How to explain this?
A big part of it is the Doomer Industrial Complex — hundreds of astroturfed organizations that have spread doomer narratives about AI.
Writer Nirit Weiss-Blatt (@DrTechlash) has analyzed this ecosystem and traced its funding to just a few Effective Altruism billionaires. Namely Dustin Moskovitz, Jaan Tallinn, Vitalik Buterin, and Sam Bankman-Fried (yes, the convicted felon).
Collectively they have donated over a billion dollars to the cause of catastrophizing AI. Those repeating the memes should understand the source.
Full article: aipanic.news/p/the-ai-existe…
Mark Li retweeted
OpenAI just published the leaderboard of the trillion-token economy.
Here it is:
1/ Duolingo – Isaac Andersen, Senior SWE
2/ OpenRouter – Alex Atallah, CEO & Co-founder
3/ Indeed – Chris Colon, Director of AI Platforms
4/ Salesforce – John Emmons, AI Leadership
5/ CodeRabbit – Harjot Gill, CEO & Co-founder
6/ iSolutionsAI – Cris Ippolite, CEO
7/ Outtake – Jiahui Jiang, Engineering
8/ Uber – Mahesh Kumar, Product Strategist (AI/ML)
9/ Ramp – Calvin Lee, Founding Engineer
10/ Abridge – Zachary Lipton, Co-founder & CTO
...and more across Shopify, Notion, Canva, Cognition, T-Mobile, JetBrains, Zendesk, Perplexity, Datadog, Mercado Libre, Genspark AI, and others.
Mark Li retweeted
ELON: THERE WILL BE NO CODE WRITING IN THE FUTURE
Remember when you had to write code? AI now spits out apps from a sentence like, “Build me a fitness tracker with push notifications.”
Syntax is dead. Abstraction won. Your job title is about to change from “engineer” to “idea whisperer.”
Sure, AI still screws up - but so do junior devs. And unlike them, AI doesn’t ask for coffee breaks.
We’ve gone from hand-coding assembly to mumbling product specs at machines that build the software for us.
Sources: Wired, MIT Technology Review
Mark Li retweeted
Just released gpt-oss: state-of-the-art open-weight language models that deliver strong real-world performance. Runs locally on a laptop!
Mark Li retweeted
Ten thoughts on AI.
balajis.com/p/ai-is-polythei…
Mark Li retweeted
every moving machine will require an inference chip, and the capability of that machine will be governed by the integrated capability of that chip
wow
The @xAI goal is 50 million in units of H100 equivalent-AI compute (but much better power-efficiency) online within 5 years

































