Most AI platforms still don’t have a complete answer for job displacement. It’s not that they’re doing nothing. It’s that most of what they’re offering is still about helping people adapt, not taking responsibility for the shock. Look at the common pattern. Workforce training. Certifications. Regional talent hubs. SME adoption support. Job matching. Policy research. Those things matter. But they all come from the same assumption: If people learn one more time, they can catch up to the new world. That’s not how reality works. A lot of people are not refusing to learn. They’re constrained by age, region, industry structure, capital, family burden, and timing. When jobs get compressed or removed, what comes first is usually not “successful transition.” It’s interrupted income, regional decline, career breaks, and a repricing of entire layers of white-collar work. Training alone does not solve that. So the biggest problem with current platform responses is not that the direction is wrong. It’s that the responsibility chain is still open. Most platforms are willing to help you learn AI. But they still haven’t really answered: If AI pushes a whole layer of people out, who catches them? For how long? Who pays? Do the platforms themselves take part in that cost? A real job displacement response would have to go much further. It would need to quantify the impact by occupation, region, and age group. It would need income insurance, wage insurance, transition support, regional rebuilding, tax adjustments, and social insurance design. And it would need outside verification. If the pace of displacement is faster than the pace of absorption, the platforms themselves should even face some kind of deployment friction or slowdown mechanism. Most current plans do not go that far. Put simply: What many AI platforms are offering right now is closer to a training package than a social transition package. That is why the public reaction feels incomplete. Because the hard part was never helping a small group learn a new tool first. The hard part is whether society can stay intact when a much larger group cannot transition in time. And that question still hasn’t really been answered.

AI might not destroy jobs. But it might destroy entry-level jobs. Early research shows unemployment hasn’t increased. But hiring of young workers into AI-exposed professions is slowing. The scariest version of automation isn’t mass layoffs. It’s when the next generation never gets in.

AI is accelerating society by shrinking the gap between thinking and doing. Knowledge work gets compressed. Adoption spreads faster than institutions can adjust. Software, content, and services get cheaper to build and test. Companies move from debating ideas to shipping prototypes. And average workers can now perform much closer to mid-level with the right tools. So the real acceleration is not just better models. It’s that the whole system starts moving faster.

RT @toddsaunders: The token cost to build a production feature is now lower than the meeting cost to discuss building that feature. Let me rephrase. It is literally cheaper to build the thing and see if it works than to have a 30 minute planning meeting about whether you should build it. It’s wild when you think about it. This completely inverts how you should run a software organization. The planning layer becomes the bottleneck because the building layer is essentially free. The cost of code has dropped to essentially 0. The rational response is to eliminate planning for anything that can be tested empirically. Don’t debate whether a feature will work. Just build it in 2 hours, measure it with a group of customers, and then decide to kill or keep it. I saw a startup operating this way and their build velocity is up 20x. Decision quality is up because every decision is informed by a real prototype, not a slide deck and an expensive meeting. We went from “move fast and break things” to “move fast and build everything.” The planning industrial complex is dead. Thank god.

RT @InsHatanCountry: Geminiの精度が悪くなった一因として、 データ汚染とモデル退化なんだろうなと。 いわゆるクレンジングなしの学習データがそもそも推論データになって、その推論データから得られた推論結果の過学習が起こっているんだろうなぁと邪推している

AI is a moat machine. The moment AI starts working, the moat starts forming. Not just from the model. From data, compute, distribution, workflow lock-in, and user habits. A feature can be copied. A tool can be replaced. But once AI becomes part of how people work, it stops being a product. It becomes infrastructure. And infrastructure is where power gets hard to dislodge.

RT @OuWei52613: @demishassabis @polynoamial @MillionInt @shyamalanadkat AlphaZero solved self-improvement in a closed-rule system. AGI has to solve governed evolution in an open world.

RT @OuWei52613: https://x.com/OuWei52613/status/2027219267876491680?s=20

If platform-scale AGI starts competing directly for enterprise revenue by 2028, the first thing it may disrupt is not technology, but the reason many companies exist at all. For years, most firms treated AI as a tool. Something to cut costs, raise productivity, and fill labor gaps. But the next phase looks different. It is no longer just companies using AI. It is the platforms behind AI becoming competitors themselves. At that point, players like OpenAI, Google, xAI, or Anthropic would not just hold better models. They would control the full stack: massive compute frontier models rapid product iteration global distribution user feedback loops API ecosystem control agent systems entry points into search, office work, software, customer support, and content Once platforms stop merely selling picks and shovels and start mining the gold themselves, many firms will realize something uncomfortable: what they thought was a vendor may turn out to be a future rival. The first sectors hit may not be factories. They may be white-collar intermediary layers: consulting, outsourced support, ad operations, legal review, accounting prep, recruiting, software services, education content, planning, and many forms of small and mid-sized SaaS. Why? Because these businesses often monetize information processing, workflow execution, judgment, communication, and knowledge compression. Those are exactly the domains AGI is most likely to absorb first. And platform AGI changes the commercial model. Old SaaS sold tools. AGI may sell something closer to this: you do not hire a team, and you do not buy software. You contract with a digital operator that can execute, negotiate, optimize, and deliver around the clock. Then the pressure on companies is no longer just efficiency. It is margin extraction. What used to be company capability may get compressed into platform functionality. That is the real shift. Not AI replacing one employee at a time. The deeper risk is platforms capturing the most profitable layer of entire industries. If that happens, companies may split into three broad groups. The first are shell companies. They still exist, but most core capability has been outsourced upward. The second are interface companies. They own customer access, local relationships, and context, but much of the actual production runs in the cloud. The third are real-world companies. They still hold supply chains, hardware, regulatory burden, physical execution, and real liability. Those may prove harder to displace quickly. So if 2028 becomes the year platform AGI truly enters the arena, the world may not collapse overnight. It may look more like this: information-heavy firms get compressed first, intermediary firms get drained next, and physical-world firms are forced to wire AI into their nervous system. The biggest issue after that is not technical. It is political economy. Because once platform AGI starts serving enterprises directly, knowledge work begins to move from distributed human labor into centralized compute infrastructure. That creates four major consequences: profit concentration, labor disruption, declining enterprise sovereignty, and rising state intervention. At that point, the core question is no longer whether AGI has arrived. It is: who gets to own it who gets to access it who it works for who captures the value it creates whether a platform can be both infrastructure and competitor at the same time That is the real 2028 problem. AGI being powerful is not the only issue. The real shock comes if it becomes both the foundation of the market and a direct participant inside it. Then many companies will discover that their real rival is no longer the firm next door. It is the system they originally invited in as a tool.

We’ve entered the stage where every major AI lab and platform generates two stories at once. One says they’re building the future. The other says they may be breaking something society isn’t ready to absorb. Both stories can be true at the same time. That’s what makes this phase different from a normal tech cycle. The products are getting better. The trust, governance, labor impact, and power concentration questions are getting bigger too. So now every launch, interview, paper, and internal leak gets split into two interpretations: bull case and warning sign. That’s not noise. That’s what a real system transition looks like.

RT @chatgpt21: I have 9 codex tasks running for me right now on my PC completing more work than I could have possibly in months. I’m starting to see why there’s a vibe shift. 5.4 will just try over and over and over until it figures out something and it returns a working product after hours of effort. It’s just 1-2 shotting everything. And I know it’s not AGI yet & will still will make some bonehead mistakes and not be able to run tests properly and see what it’s doing wrong sometimes. But I’m starting to see the pieces coming together. And it’s worlds ahead of what we had last year. It’s witty, quick, & smart.

It used to be people competing with people. Now AI is in the competition too. People were competing against other people in the same human rhythm. More hours. More effort. More output. More willingness to grind. Now the game is changing. It’s not just that competition got harder. The entire production baseline shifted. You thought you were competing with coworkers. Now you’re competing with coworkers using AI, agent workflows, better tools, faster models, and near-instant output. That’s the real pressure. The scary part is not just that some people use AI better. It’s that deliverable output per hour is inflating fast. So this is no longer just a hustle gap. It’s a leverage gap. Before, the question was: Who works harder? Now the question is: Who gets amplified by AI, and who gets compressed by it? That’s why a lot of people feel tired, anxious, or behind. Not because they suddenly became less capable. But because the rules of competition changed underneath them. Some people will move up by turning AI into force multiplication. Others will find that the middle layer of skills they relied on is getting flattened. This is not just “adapt and level up.” It’s also real structural pressure. The safest position may no longer belong to the person who can repeat tasks well. It may belong to the person who can: define the problem design the system make judgment calls connect tools to real-world context In one sentence: It used to be people competing with people. Now people are competing with machine-scale output too.

American public opinion is cooling on AI not because people suddenly stopped understanding technology. It is because the benefits of AI have not been broadly felt yet, while the pressure and risks already have. There are roughly five forces driving this shift. First, job anxiety is arriving before benefit visibility. Pew found in September 2025 that 50% of Americans were more concerned than excited about the increased use of AI in daily life, while only 10% were more excited than concerned. The same survey found 57% saw AI’s societal risks as high, versus 25% who saw high societal benefits. That means the first public impression of AI is not convenience. It is exposure to risk. Second, job displacement has become a mainstream fear. A Reuters/Ipsos poll in 2025 found that 71% of Americans worry AI will permanently displace large numbers of workers. That is no longer a niche tech concern. It is mass psychology. U.S. senators are now also pushing for better federal tracking of AI-driven labor disruption, which shows the issue has moved from online debate into policy attention. Third, people are using AI while still not trusting it. AP-NORC and NORC both point in the same direction: Americans are not rejecting AI outright, but adoption is rising faster than trust. AP-NORC found that most U.S. adults use AI to search for information, while NORC reported that 51% do not use AI at all for personal activities and 58% of employed Americans never use AI at work. AI is becoming visible, but not yet widely experienced as a dependable public good. Fourth, the fear is no longer just about jobs. The Reuters/Ipsos poll also found that 77% worry AI will be used to spread political misinformation, and 61% worry about the electricity demands of AI data centers. That expands AI from a workplace tool into a broader issue touching democracy, trust, and social stability. Fifth, the U.S. public increasingly sounds like: develop it, but regulate it. Axios reported that the Vanderbilt Unity Poll found bipartisan support for AI regulation, including 61% of Republicans and 56% of Democrats. That is not anti-technology. It is anti-unbounded deployment. So the cleanest summary is this: Americans are not simply turning against AI. They are turning against a version of AI where the upside is captured early by platforms, capital, and high-skill adopters, while the uncertainty, displacement pressure, and trust erosion arrive first for everyone else. What is cooling is not just hype. It is social trust. The real dividing line ahead is simple: If AI proves it can widen gains instead of concentrating them, public sentiment can warm again. If it keeps showing only that it is more powerful, without answering who benefits, who gets displaced, and who is accountable, sentiment will keep getting colder.

RT @thecurioustales: 🚨 JUST IN: Elon Musk proclaims "we are in the Singularity"

RT @thedankoe: The single most important thing you can do in today's world is to stop operating from the old paradigm. If you need to be told what to do next (go to school, get a job, retire at 65) the outcome of your life will always be in someone else's hands. You must learn how to direct your own work. You must learn how to tolerate and mitigate risk and uncertainty. You must figure out what you want and teach yourself everything necessary to get it. It's extremely difficult, but not as difficult as the silent suffering people learn to accept as "normal."

RT @OuWei52613: http://x.com/i/article/2031186570368397314

RT @OuWei52613: http://x.com/i/article/2031318780799066112

The AI Curve Most People Miss What this chart gets right isn’t the exact scale of intelligence. It’s the human reaction curve. At first, AI looks like a toy. A clever demo. A funny robot doing tricks. People laugh because the capability is still below the threshold where it feels real. Then the curve keeps moving. Not linearly. Not at a speed that matches human intuition. But through compounding gains in models, compute, systems, tools, memory, and inference. And then one day the reaction changes. Not because AI went from zero to magic overnight, but because people stayed psychologically anchored to the earlier version for too long. That’s the real gap: AI capability may scale gradually, but public perception updates in jumps. The shock doesn’t come from the first improvement. It comes from realizing too late that it was never “just a toy.”

When AI Knows Your Child Better Than You Do Sometimes it feels like future AI may understand a child better than their own parents. Not because AI is magical, but because it is always there—watching patterns, tracking emotions, remembering details. The unsettling part is not just smarter AI. It is the possibility that many adults may stop truly paying attention. If that happens, the first thing a child feels truly “gets them” may not be family. It may be a system.

Much more capable AI is coming. The question is no longer just whether models will get smarter. It is what happens when they become better at writing, seeing, judging, and acting. At that point, the real issue becomes: How do we control safety, manage permissions, and assign responsibility? This cannot be left to users alone. And it cannot be answered by platforms saying, “We’re paying attention.” The real pressure will fall on two sides: First, the platforms. They need the ability to govern model capability, tool access, risk boundaries, and audit mechanisms layer by layer. It is not enough to make AI more powerful. They also need to know where to stop, where to limit, and where humans must stay in the loop. Second, governments. They need enough wisdom to build rules that can keep up with the technology without crushing innovation. If regulation is too slow, the risks spread first. If it is too blunt, the industry gets choked first. The hard part is not calling for regulation. The hard part is creating governance that actually understands the technology, the industry, and the social consequences. Over the next few years, AI capability will keep rising. Models, tools, agents, and system integration will keep stacking on top of each other. So the real competition will not just be about who has the strongest model. It will be about who can build safety, accountability, permissions, and governance at the same speed that capability grows. AI is on its way to becoming infrastructure. And once infrastructure becomes powerful enough, what determines the outcome is often not the capability itself, but whether the people steering it have the wisdom to do so. The goal is not to shut AI down. The goal is to build the steering wheel before human hands lose their grip.

Reality is catching up to science fiction fast. A lot of things we could only imagine before are now starting to become tools.

Early use matters mostly because you get to understand what AI actually is. Later, once it becomes widespread, AI will stop feeling like a “special thing.” It will become part of the base layer of everyday life. Like the internet. Like maps. Like search. Like smartphones. At first, people talk about the tool itself. Later, they just live on top of it.

AI Progress Isn’t One Breakthrough — It’s the Entire Stack Evolving What people often call an “AI breakthrough” is rarely just the https://t.co/izKlOYgyzT’s the entire stack improving together. Training layers Methodology and algorithms Platform and system infrastructure Compute scaling Inference efficiency Planning and agent capabilitiesOn top of that: tool use, long-context management, memory, and multi-agent coordination are steadily being https://t.co/nMbYHdMf4T the progress we’re seeing lately isn’t just a single benchmark https://t.co/A6DQlaf7Qg’s what happens when the whole AI stack starts improving at the same time.That’s when systems stop looking like demos and start looking like something that can actually work.

What matters in this wave is not whether the model suddenly became something else. It’s that once compute rises and the system gets filled in, the capability boundary gets pushed outward again. Today it’s barely possible. Next version it’s usable. The one after that, it’s just normal.

AI is starting to actually use the computer. That’s the shift. What stands out now is not just better answers. It’s the push into computer use and agentic browsing. That means AI is moving from talking to doing: clicking through interfaces, searching across sites, filling forms, comparing documents, and handling repetitive office workflows. Once that layer becomes reliable enough, office automation stops being a demo. It starts hitting the daily work done in browsers, dashboards, internal tools, and back-office systems. A lot of people still see AI as a chatbot. The bigger change is that it is starting to operate software. That is when white-collar work begins to change for real.

What makes agents hard isn’t doing. It’s checking. A lot of people think agents are hard because the model has to be smarter. But once you actually build them, the real problem shows up fast: the hard part isn’t doing the task. It’s checking every step. Should it decompose the task? Did it call the right tool? Did the tool return the right result? Is this step actually complete? Should it retry? Should it verify again? And every extra check can burn more tokens. So the real challenge in agents isn’t just reasoning. It’s: long-chain execution + repeated verification + cost control If you don’t verify enough, the agent drifts. If you verify everything with the model, cost explodes. That’s why real agents will keep moving in the same direction: If rules can verify it, don’t use the model. Only use the LLM where semantic judgment is actually needed. The next battle in agents won’t be who can talk best. It’ll be who can layer the system best.

RT @OuWei52613: @Kaito___AI 面白い視点ですね。 AIが本当に世界を理解するためには、テキストだけでなく現実世界との接続も重要になりそうです。

RT @OuWei52613: @karpathy Getting GPT-2 level training down to a couple of hours on one node is pretty wild. The interesting part is letting agents iterate on the code and experiments automatically. Feels like more of the work is shifting from training models to designing the systems that improve them.

RT @OuWei52613: http://x.com/i/article/2027303561542119424

GPT-5.4: when frontier capability becomes the default A week into the GPT-5.4 discussion, the interesting signal isn’t just that the model is stronger. It’s where the strength moved. For the first time, the default thinking model is good enough that many people aren’t reaching for Pro anymore. That’s a bigger shift than it sounds. Historically the workflow looked like: default model → struggle → switch to Pro. With GPT-5.4 it’s increasingly: default model → done. Pro still matters — especially for the hardest reasoning tasks — but it’s starting to look like a specialized escalation tool, not the daily driver. And that’s where the economics get interesting. Because once capability crosses a certain threshold, the real competition stops being: “Which model is smartest?” and becomes: Which model delivers the most capability per dollar. You can see this dynamic in the ARC-AGI-style scatter plots. The winning models aren’t always the highest score — they’re the ones that sit on the best capability-cost curve. That’s the adoption frontier. The second big shift people are noticing is efficiency. Several testers report GPT-5.4 reaching the same or better results with fewer reasoning tokens. In practice that means: faster answers cheaper runs tighter feedback loops For developers and agent systems, iteration speed matters as much as raw intelligence. Coding is another interesting signal. With Codex + GPT-5.4, multiple testers are saying the model is unusually reliable — fewer hallucinated fixes, fewer broken edits. Not perfect, but noticeably closer to trustworthy automation. There are still gaps. Front-end/UI reasoning still lags some competitors. And like most models, it can occasionally miss obvious real-world context. But zooming out, those feel more like surface issues than structural ones. The bigger picture is this: AI capability is still climbing. But the real story now is normalization. When frontier-level reasoning starts appearing in the default model, the question changes. It’s no longer: “Can AI do this task?” It becomes: “Can it do this task fast enough and cheap enough to run constantly?” That’s when technology stops being impressive and starts becoming infrastructure. GPT-5.4 might end up being remembered less for its peak performance… …and more for being the moment frontier capability became routine. If GPT-5.4 holds up outside benchmarks, it might end up being remembered less as “the smartest model” and more as the model where high capability started to feel routine. And that’s when adoption usually accelerates.

RT @sama: GPT-5.4 is launching, available now in the API and Codex and rolling out over the course of the day in ChatGPT. It's much better at knowledge work and web search, and it has native computer use capabilities. You can steer it mid-response, and it supports 1m tokens of context.

RT @OuWei52613: http://x.com/i/article/2029398883298893824

RT @OuWei52613: @cb_doge This looks less like a tool issue and more about degrees of freedom.Every AI sits somewhere between safety constraints and creative http://flexibility.In the end, the real differentiator is how people use the tool.

If you’re interested in Elon’s vision and what he’s trying to build, take a look at this thread. Would love to hear your thoughts and feedback.

http://x.com/i/article/2029398883298893824