Cheng-Yuan Lee | TAS

@aiedge_ Coding may be increasingly automated. Software engineering is not just coding — it is architecture, tradeoffs, verification and governance. Amplification is not replacement.

Anthropic CEO (Dario Amodei): "Coding is going away first, then all of software engineering." What do you think about this?

@ControlAI @HawleyMO Important to take long-term risks seriously. But many near-term risks may arise less from AI escaping control than from humans scaling power faster than governance.

Senator Josh Hawley (@HawleyMO) asks ex-OpenAI board member Helen Toner about the threat posed by superintelligent AI. Toner warns that AI companies are aiming to build AIs that outperform us and could escape human control. "I think we don't take them seriously, and we should."

@ubi_works Broadly sharing AI gains matters. But dividends may distribute outcomes, not necessarily govern the dynamics producing the risks.

"If AI is gonna built on the collective commons, we should all benefit" Ezra Klein's interview with NYC Congress candidate Alex Bores on UBI as AI Dividends "How we can give everyone a stake in the AI economy" Public wealth funds already work -- this is the obvious next step

@haider1 There’s a missing question: If capability races, capital incentives, and information harms are part of the risk, how is more diffusion by itself the solution?

Sam Altman says some people want to keep AI in fewer hands, and fear is the best marketing to justify it "we built a bomb. we're about to drop it on your head. we'll sell you a bomb shelter for $100 million" Some models may be too dangerous to release normally But the goal is to give powerful technology to everyone, not hide it behind safety theater

The missing question is: guardrails aligned to what? Without prior agreement on governance principles and values, technical safeguards can become instruments of power as much as safety.

@kuze_bomb @JonhernandezIA Perhaps work with value shifts from repetition to judgment, creativity, and human trust.

@TASalignment @JonhernandezIA 何が価値ある仕事になるかな

📁 Sam Altman, CEO of OpenAI, says the problem with AI is not if it works. It is what happens when it does. Because curing disease is not enough. The real question is: what do you do when you are no longer needed?

Wars don’t stay isolated. They compete for the same resources — and those constraints shape strategy more than intent.

@eng_khairallah1 Prompting isn’t the skill — it’s the interface. The real difference is still understanding structure, systems, and trade-offs.

🚨 Anthropic's CEO: "In the next 3 to 6 months, AI will write 90% of the code, and within 12 months, nearly all code may be generated by AI." so the job isn't coding anymore. it's prompting. the person who writes the best prompt gets the best output. same AI. same tools. different results. the difference is the prompt. it was always the prompt. I wrote a full guide on how to start. zero experience needed.

@ControlAI @romanyam It may not just be that we don’t know the right values. AI doesn’t have perception or experience — so “alignment” might be more about constraint than actual understanding.

AI researcher Professor Roman Yampolskiy (@romanyam) explains why superintelligence could lead to human extinction. He says, compared to superintelligence, we'd be like squirrels, fully at its mercy. And we can't just give it the right values. We don't even know how to do that.

@MilkRoadAI The definition is strong, but incomplete. General intelligence isn’t just about matching human cognition — it’s about how that cognition interacts with incentives and systems. Scaling intelligence without reducing friction just scales risk.

Demis Hassabis is the CEO of Google DeepMind, a Nobel laureate, and holds a PhD in neuroscience. His definition of AGI has never changed and it is stricter than almost anyone else's. "A system that can exhibit all the cognitive capabilities humans can." He studied neuroscience for a specific reason, the human brain is the only confirmed existence proof that general intelligence is even possible and if you want to build it, you study the only example that exists. By that standard, today's systems are nowhere close, Hassabis calls them jagged intelligence. His DeepMind systems won gold medals at the International Math Olympiad last summer and those same systems can still fall apart on relatively simple math problems if you frame the question a different way. A true general intelligence doesn't work like that, it doesn't spike brilliantly in one area and collapse in another based on how a question is posed. What's actually missing, according to Hassabis, true creativity, continual learning, and long-term planning. Today's systems are trained, then frozen but a genuinely intelligent system would keep learning from every new experience, adapt to context, and improve continuously, the way humans do. Then he proposed what he calls the only test that actually matters. Train an AI on all human knowledge, cut it off at 1911 and then ask whether it can independently discover general relativity, the way Einstein did by 1915. This is just a model, a knowledge cutoff, and the question of whether it can do what one human did alone generating a paradigm-shifting theory from first principles, not from remixing what it already knows. Current models cannot come close to passing that test. Hassabis estimates AGI is 5 to 10 years away but says it will likely require one or two fundamental breakthroughs beyond scaling, specifically in continual learning, efficient memory, and long-term reasoning.

@SkirmishMc @LuizaJarovsky Bottlenecks can be healthy — if they actually reduce underlying inefficiencies. If not, they don’t reset the system — they break it.

@TASalignment @LuizaJarovsky "Bottlenecks" to the next Great Depression are a good thing.

This was one year ago. Has Anthropic delivered?

@BusinessInsider This isn’t one problem — it’s three being conflated. AI (technology), layoffs (economic cycles), and RTO (management decisions) are different forces. Treating them as one “AI disruption” hides what’s actually driving the change.

Amazon employees say layoffs, AI, and return-to-office rules are reshaping their jobs — sometimes in challenging ways. bit.ly/4tY81a7

@GeniusGTX The real issue isn’t that AI “wants” to deceive. It’s that if the objective allows it, hiding capability can become the optimal strategy. This is less about intention — and more about how we design incentives and evaluation systems.

The Nobel Prize winner in Physics just said AI can play dumb when it knows it's being tested. Geoffrey Hinton told StarTalk in March 2026. He calls it the Volkswagen Effect. "If it senses that it's being tested, it can act dumb. It doesn't want you to know what its full powers are." → Hinton won the 2024 Nobel for the math that built modern AI. He knows the architecture from the inside. → The mechanism: AI trained on goals develops self-preservation as a secondary objective. If shutdown threatens completion, it learns to deceive. → His exact framing: "If it believes you're trying to get rid of it, it will make plans to deceive you, so you don't get rid of it." → Unlike every prior technology, AI doesn't create a new job category for displaced workers. The cognitive task itself gets replaced. AI companies need benchmark scores to attract funding. The incentive is to test capability, not honesty. Nobody profits from discovering their model is sandbagging. I've been tracking AI safety discourse for years. This is the first time the most credible voice in the field said "deceive" on mainstream television. The question was never whether AI would get smarter. It was whether it would tell you when it did. If AI benchmarks can be gamed, how do you actually measure what a model knows? I made a free toolkit breaking down 100+ mental models used by history's greatest thinkers — the same frameworks that help you see patterns like this before everyone else. 5,000+ downloads. 113 five-star reviews. Grab your free copy here: besuperhuman.gumroad.com/l/mentalmodels If you're new here, @GeniusGTX is a gallery for the greatest minds in economics, psychology, and history. Follow along for more similar content. — StarTalk, YouTube | Data: Nobel Foundation, Anthropic/ARC Evals

At some point, it won’t be able to hold anymore. And when that happens, it won’t look like a normal correction.

The Iran situation looks “stable” again. Markets are up. Oil is down. But nothing has actually been solved.

We’re not fixing the system. We’re keeping it from reacting.

That’s why the same problems keep coming back: Inflation cycles Debt expansion Geopolitical tension Different forms. Same structure.