Personal update: I’ve joined @Meta to build AI products.

While SWE agents have transformed software engineering, agents in most other complex systems are still early. Most people haven’t yet felt the full power of AI agents.

Meta is well positioned to change that.

Time to build.

Great post on some of the dynamics to think through for the future competitive advantage in world when AI models are shared amongst firms and packing so much for the intelligence of that industry.

This is going to become a core question for companies and the economy broadly over the next decade and beyond. If AI is trained on the best datasets in every single industry - like law, finance, healthcare, or life sciences - then how do you compete and differentiate in the future?

This is a great open question that I don’t think is perfectly knowable right now because of how fast AI progress is happening. But ultimately it stands to reason that if intelligence is abundant and broadly available to anyone in a field, then the companies that effectively use it the best and against a set of data and knowledge that grows in value over time, will be in a strong position.

There’s a huge reinforcing loop between the intelligence from models, a company’s own data, the connection of that data and AI in their workflows, and how employees ultimately interact with that system to create value. There’s no obvious point where this will become uniform across all companies in a vertical because each company will approach this in a different way, just as they already do with their talent and workflows. If anything, there will be compounding returns to those that do this best that accelerate their advantage over time.

Overall, super interesting question to see how this plays out over time.

GPT-5.6 is now out. We've been evaluating the model family on the Box AI Complex Work eval, which tests the model with the Box AI Agent on a variety of extremely hard tasks using enterprise document sets.

Sol is a big step up from GPT-5.5, especially on complex data-oriented tasks that require deep reasoning and analysis, and the gains concentrate exactly where enterprise work is hardest. Here are a few examples that we saw across our tests:

* Financial Services (76% vs 71%): On a multi-year projection, Sol anchored to the correct opening balance sheet date rather than assuming a clean January 1 start, then carried revenue, earnings, and interest through to the right figures year over year where one early wrong assumption compounds through every downstream cell.

* Healthcare (58% vs 46%): On a critical-care case review, Sol identified the correct diagnosis and intervention and avoided the dangerous misstep of ordering imaging before the time-critical procedure, a trap GPT-5.5 walked into.

* Public Sector (74% vs 63%): Handed a class's raw gradebook and a new grading directive, Sol mapped each assignment to the right weight bucket, treating homework as zero-weight practice per the directive. It recomputed every student's grade to within a tenth of a percent, where GPT-5.5 drifted partway through.

* Life Sciences (60% vs 51%): Across four separate compound datasets, Sol intersected the ranked target lists exactly (case-sensitive, no shortcuts) to find the biological targets common to all four, catching the shared targets GPT-5.5 missed.

Sol reasons from the source definitions and checks the documents rather than taking them at face value and it's most reliable exactly where the numbers drive real decisions. This will be huge for enterprise agents using unstructured enterprise data.

GPT-5.6 will be available to customers shortly within the Box AI Studio for building custom agents with.

Meta Muse Spark 1.1 (early access was called Hornbill) turned out to be really good on my OpenClaw

Way to go @alexandr_wang https://t.co/HK3Z9TrJA1

While AI is making coding less rigid, we’re making the runtime more rigid.

I’m noticing our infra teams writing formal specs for the first time. More deterministic systems. More resilient infrastructure.

The faster you want to move, the more solid the ground beneath you has to be.

It’s fantastic to see how dynamic the LLM market has become in just a short period.

Just 6 months ago VCs suffered Anthropic psychosis and convinced themselves that it was going to be a monopoly.

They will keep making great models, but so will the others, including new entrants

If you haven’t tried Fable in Replit… https://t.co/AzAHNxDPU5

Me explaining this week’s model release situation to my wife..

"babe listen. this matters. GPT-5.6 dropped in three flavors. Sol, Terra, and Luna. Sol is the smart one. It had to clear government safety evals before release, like getting paroled. Luna is the cheap one. Terra exists.

Grok 4.5 launched the day BEFORE GPT-5.6. On purpose. So it could have 24 hours of being frontier-adjacent. Elon calls it Opus-class, which is a polite way of saying it competes with Anthropic's previous model. But it's $2 per million tokens, so nobody will care.

Speaking of Anthropic. Fable 5 just came back online after the government suspended it for three weeks. The best coding model on earth was literally grounded. It's back now, and the rumored GPT-6 is being retrained specifically to beat it.

Sonnet 5 also shipped last week. Near-Opus performance at $2 per million. It's the new free default. Meaning the free tier is now better than what enterprises paid six figures for in January.

Meituan open sourced LongCat-2.0. That's a 1.6 TRILLION parameter model. MIT licensed. From a Chinese FOOD DELIVERY company. Your DoorDash equivalent is now a frontier lab. Let that one sit.

ByteDance launched Seedream 5 Pro. Almost as good as GPT Image 2 at 4x cheaper, because ByteDance's entire corporate strategy is 'same thing but cheaper.' It beat Meta's new Muse Image, which trains on your public Instagram photos unless you opt out. Your 2014 gym mirror selfie is now foundational research.

OpenAI also shipped GPT-Live, a voice model that goes 'mhmm' and 'got it' WHILE you're still talking. We finally taught AI to pretend to listen. Full parity with husbands achieved.

And Ollama raised $65M so you can run the open source versions of all of this on your own laptop, which is great for privacy and also heating your apartment.

And btw it’s just Thursday"

As a former electrical engineer who got in trouble for doing this in college, I love this so much ⚡️ https://t.co/dA1emRNWir

whoever does AEO for @resend needs to get a raise, all the leading frontier models keep trying to use resend for emails even when i have existing transactional infrastructure already set up https://t.co/vxJoVSF9ju

u guys were clowning on @greptile but turns out they were just the inspo for @openai to go so, so much harder https://t.co/cIarQMR3LD

more grist for the mootools mafia lore

https://t.co/m2ZCSJlLrv

For the curious, these were the east harlem and east village gas explosions in 2014 and 2015 respectively. But those are 2 of the 19 events in the wikipedia list of structural collapses in New York, which seems to go back to the 1800s: https://t.co/Yiu8SNLdA0

When I was living in New York, one building in my friend's neighborhood collapsed. The next year a building near me collapsed. This left me with the impression that buildings did just collapse somewhat regularly in New York. Turns out that's not actually the case, which is good.

Watching this made me remember how insanely great the 2022 World Cup final was. One of the best of all time.

Maybe a rematch this year? https://t.co/4NZj9lXWrx

Some praise and feedback about OpenAI's launches:

1. More than any other lab, OpenAI has the opportunity to make working with agents mainstream. ChatGPT with images, live voice, browser/computer use, and plugins for all your favorite apps is the closest thing to working with a super-capable coworker who can learn and do any white collar work.

2. I didn't get to test GPT 5.6 weeks early, but after a day of testing, I think my best compliment is: "It's got that dog in it."

It basically never gives up, no matter how complicated a task you give it, and is very reliable thanks to browser use and the other features above.

It does seem to burn more tokens than GPT 5.5, although maybe I don't have the right settings. I'm using Sol on High.

3. GPT Live is arguably a more important launch than today's updates for the masses.

We've been talking to each other using voice for 100K+ years, while the keyboard and mouse weren't introduced until the 1950s/60s. Maybe voice will soon make these two tools look like old computer punch cards.

Also, the best thing about GPT Live is that it lets me pick the intelligence level. I care about latency, but I care more about talking to someone who's smart and gets me.

Speaking of which, it'd be cool if I could have a live conversation with ChatGPT voice and see the tasks it's doing for me.

Ok, now for some constructive feedback:

4. I think the ChatGPT Work vs. Codex thing is confusing. It raises questions like, "So developers aren't working?" and "What if I'm using it to plan my vacation, is that work?"

I think the whole thing should just be called ChatGPT, Codex, or ChatGPT Codex and there should be no tabs or toggles. IMO, Codex is not that complicated for a normie to understand and use. You're still chatting, except now it can do stuff.

The sooner this is all unified, the better.

5. I'm very confused about when I should use Sol, Terra, or Luna, and whether I should set effort to Light, Medium, High, etc.

There's nothing in the UI that walks through what to use when. I understand power users might want these options, but regular people will just get confused.

6. Another thing that's confusing is tasks vs. chat. Ideally, your past chats should still be there in the left nav, except now those chats can do something with plugins. Normies will not understand why their chat history is suddenly harder to find.

7. When I was a PM, I loved building with a community of trusted users who gave me feedback along the way before general availability. But I also tried to make it transparent to users what qualifications were needed to join this community. I don't mind that OpenAI is doing the former, but the latter is not clear at all.

Thanks for listening. Codex has changed how I work and for that I'm very grateful.

Ugh France is way too stacked

PS: open models are about to get exorbitantly fast. Watch this space

It’s model release week. I suspect Meta Spark 1.1, Grok 4.5, and GLM 5.2 will significantly displace token market share.

Most agentic tasks require reasonably high intelligence at fast speeds.

It’s a good time to be using ▲ AI Gateway! https://t.co/aK4FSpnuhj

X is the arena. Always has been

one of the core skills of agentic coding is reducing your unknowns https://t.co/XkyGDHBZqn

Enjoy more Fable! https://t.co/DHXH1stWma

Decent chance France and Norway could play again in the final, which would be an extremely rare occurrence (been over 60 years since two teams met in both the group stage and the final of the same men's World Cup) and fascinating (how truly game-changing is Haaland?)

Hang it in the Louvre. https://t.co/4XJSCK7K1a

X: “Paris will be in flames after the France-Morocco match”

Paris after the match: https://t.co/UzFZ3jLFzj

ChatGPT Codex Adeptus Astartes

Two bros talking at the screen with a big dollar number graphic who cares?

Why do people make flashy videos talking about how much money they raised? In what way does that benefit your business?