Meta Chief AI Officer Alexandr Wang on why your product can be better and still lose, because perception is more real than reality: Most founders assume the best product wins. Build something that works, prove it with data, and customers will see it. Wang says that assumption breaks down when you sell to large organisations. "Probably a lot of the companies that you all have worked with, they're very data-driven companies. It feels like the truth makes its way... everybody serves a shared sense of reality and whatnot. That's not true at most large companies and also not true within the government, unfortunately." So what do they run on instead? Perception. "At a lot of large customers, perception is more real than reality. The reality is just so ugly most of the time that very rarely do people actually confront reality, and most of the time they just sort of choose to believe the perceptions that they live in." That is how a better product loses. If buyers aren't judging reality, being better in reality isn't enough. The deal goes to whoever owns the perception. Wang spells out what this means for anyone building or selling to enterprises: "If you end up doing enterprise sales or you end up building enterprise products... just as much as your job is to improve the reality, it is to shape the perception." The product still matters, but by his reckoning, it's only half the job. One company he singles out for mastering the other half is Palantir: "One of Palantir's superpowers is that they shape the perception better than most other technology companies do, because I think they view themselves... it's like a combination of a sort of acting troupe combined with a software company." And @alexandr_wang means it literally: "They literally give an acting book to all of the new hires, or they did for a very long time." A software company that trained its people in performance. That tells you how seriously Palantir took the perception half of the job.

You're using codex cloud wrong. this should be the way: - the codex / claude code "client" should not only be on you're laptop - it should also be on a cloud (you can use dot to do this) like a bastion server Why? if you're using multiple agents across providers to communicate & orchestrate each other, the client should be active to receive other agents requests. You can also try to install the client ON the agent's cloud, (install cc on codex cloud) but getting auth on every cloud env is kind of a hassle.

Watch GPT-6 Astra synthesize a molecule. This facility is a physical harness for AI – it lets Astra work on all parts of R&D, including those in the real world. Unlike math, science is bottlenecked on verification loops. This is the infrastructure needed to compress them.

We gave GPT-6 Astra a jet factory. Then connected Jev for typed, verifiable decisions. Astra understands the production system. Jev reduces each critical decision to a clear yes/no probability. Deterministic code applies the approved change. The result: an AI-native factory that can identify bottlenecks, inspect processes and improve production in real time. Try Jet Plant II: airsup.ai/factory/jet-plant-…

We built high-throughput materials labs in Menlo Park to create a loop between experiments and models. The labs generate fresh data, the models learn from it, and then help us decide what to try next. Using only 1,300 H200s, plus months of our experimental data, we mid-trained and RL’d an open-source model to surpass GPT-6 Astra on our analysis benchmark. We call it Neon. This is real footage from our lab. We’re focusing first on hard problems in materials science, including superconductors, magnets, and semiconductor materials. Read our blog posts below.

nothing feels better to prove your doubters wrong with your product that was already thought out 10 years ago

In the end, it all comes down to data & distribution.

A lot of people are missing Terence Tao’s point and thinking “mathematicians are upset that AI is better than them.” That’s not what he’s saying, and some people are forgetting that Tao is one of the most AI-pilled mathematicians out there. His point is that when people work on discovering something, along the way they invent new concepts. Those concepts later become useful far beyond the original goal, and enables further inventions. Finding a solution does matter, but the intermediate idea is often what makes the field richer, because other people can share it and build the next thing from it. In tech, we can use the analogy of collaborative software. We started with algorithms for merging changes in a Word document, and evolved that to concepts about versions, diffs, and merges, and later to real-time collaboration tools like Git, Google Docs, and Figma. Humans built upon these concepts and developed more powerful solutions. Terence’s worry is that a machine automating a solution robs the field of the value of developing the intermediate discoveries in the pursuit of larger discoveries. When automating a solution, the intermediate discoveries and invention of concepts can be buried or completely hidden in the black box. We don’t learn from them to build the next thing; it’s like we never made the invention of collaborative document editing and thus could not have the conceptual understanding to invent the next version – and since it’s hidden, we also don’t socialize them to allow other people to invent, too, a core tenet of collective discovery. So then, in both code and math, this leads to the atrophy of development of concepts in the field. In other words: pure ‘solution extraction’ that hides the process of discovery can leave the field with a checked-off theorem but little new insight or new questions to pursue. And it might prevent us from understanding a field deeper. I am seeing, first-hand, that atrophying of skills in software development. We push buttons and get solutions. There is much less incentive to develop new concepts and human skill. The bet most software companies are making is that LLMs are so effective in writing code that you’re still shipping overwhelmingly more value even with human skill atrophy, and it’s the right bet IMO. However, much of the software industry is built upon building things, not necessarily novel invention and research. In such an environment, you can say that you accept some atrophying of conceptual invention and human skill for more output. On the other hand, sectors like math and pure sciences that are focused on invention and insight might be the hardest hit by this. Practical/applied sciences might fall somewhere in the middle. An Alzheimer’s cure, room-temperature semiconductor, or highly effective carbon capture solution are far too valuable to sandbag and say only humans can do that to develop concepts in the ‘proper’ way. The outcome matters too much to treat the preservation of concept invention as the highest goal. Even there, though, hidden intermediates can slow the next breakthrough if nobody can see how the first one actually worked. So the question is not “is AI allowed to solve hard problems?” It is “in this field (math, science, tech, etc.), is the answer itself the main point, or are the concepts and abstractions we use to get there also the thing we need to maintain?”​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​ In pure math, there’s an argument that the intermediates are often more useful than the solution, and atrophy in concept development is highly detrimental to the field. Solving Navier–Stokes, contrary to what some people claim, has little practical application, and pure math might be one of those fields where just finding a solution isn’t the entire point, and can actually be contrary to the field, which is what Tao is worried about.

Stopping is not an option. Distributing it as widely as possible and as quickly as possible is the best choice we've got. Better everyone have it than only the tech companies who have been manipulating you for years.

I am pleased to see that Ant’s new Claude Code has a version of background computer use on par with the version from Codex from last May this year. Nice work! Shipping great features first turns out to be a great way to encourage other labs to ship too. We will continue to do this until other labs pay more attention to shipping. This is good for everyone and there is a lot of room left to go! We solved computer use in practice for GPT models about four months ago. But computer use is a significantly awesome thing no matter what model you use, and it is important that the industry similarly spends more effort to train their models to be great at computer use, among other elements of model capabilities. As models become more capable and central to businesses and economies, the value of computer use only increases. We should be taking it seriously, and doing the right thing for our customers and the world.

I had my “touch of AGI” moment yesterday. I recorded a video of myself talking through a problem: how to improve maintenance of tangled hair in the shower drain. It was a ~4 min video where I explained the problem and also used a digital caliper to take measurements. I asked GPT 6 Astra (Ultra) to come up with a fix for the problem. It took the video, extracted my voice, and matched the video frames to where I measured the parts. It came up with an idea for making maintenance and cleaning easier. I got an interactive sketch so I understand what it is talking about. I told it to use computer use in Fusion 360. I watched in real time as it designed the part, making it fully parameterized. While watching, I learned a couple of tricks, like how making a design read-only frees up one of the 10 free editable designs you get 😅 Anyway, it designed it and it looked solid. I told it to add fillets. I 3D printed it, tried it, and it fit perfectly on the first try. Even though it wasn’t super complex, it solved a real-world problem, probably faster than it would have taken me to wrestle with Fusion and get what I wanted.

Do we want “good code” or do we want a "good product"? If what you actually care about is making a good product, AI is already good enough to write pretty solid code even if you don't really know how to write good code yourself. And that's obviously only gonna get better as the models get better. Sometimes you do need to understand the code though. Maybe the code itself is the product, like with an open source project. Or maybe you just need to be able to catch it when the AI screws something up and tell whether the code is actually correct. But I see people take this as “so learning programming or CS is pointless now,” and that's not what I'm saying at all. That's a completely different thing. A lot of people talk about AI like it's democratizing knowledge. I think it's more like an amplifier. The more you know, the better you are at telling the AI what you actually want, and the more likely you are to get a good result. It's kinda like having really smart employees at a company with a dumb CEO. Smart employees don't magically fix that. Knowing what to ask for, and being able to clearly tell people what you want, still takes a lot of knowledge and effort.

GPT-6 Astra设计FPV无人机飞控 PCB全过程记录，从电路原理图到 PCB 布局、走线、元器件摆放，都是Codex里的GPT-6在 KiCad 里直接操作，我可以看到它一步步画出来。 画完后它自己运行 ERC、DRC，检查问题，再修改电路和布局。 最后完成了 F405 六层飞控板的设计，还导出了 Gerber、钻孔、BOM、贴装坐标和 STEP 文件，能直接打开 3D 看整块板子的效果。 整个设计过程，总耗时3 小时 13 分 56 秒，总token消耗量约为51,900,896个（约5200万个），模型思考强度为高。 目前还没打样，实际能不能稳定工作，还要上板测试。但从描述需求到画出整块 PCB，这个过程已经挺让我惊叹了。 #AI #PCB #KiCad #Codex #GPT6 #Astra

Software engineers, it's time to go hardware.

Called it.