High Signal AI @HighSignal_AI

This fake Kanye x Elon conversation built from real interviews using AI is the most unhinged thing you'll watch today. Source: THE TALK

Most people talk about AI like it appeared overnight. It didn’t. What we call “AI” today is the result of decades of stacked breakthroughs, With each layer building on the limitations of the one before it. Sharing great insights from Luís Rodrigues. Here’s the breakdown of the AI layers: 1. Classical AI Rule-based systems. “If this, then that.” No learning involved. No flexibility. Just predefined instructions executed exactly as written. 2. Machine Learning Shift from rules to data. Feed enough examples, and the system identifies patterns on its own. It uses past data to predict outcomes. Data over hard coding. 3. Neural Networks Inspired by how the brain works. Connected layers of artificial neurons process complex inputs like images, audio, and text. This is where machines start handling real-world ambiguity. 4. Deep Learning Neural networks at scale. Deeper architectures unlocked real performance in language and vision. This is the backbone behind modern AI breakthroughs. 5. Generative AI From understanding to creating. Models now generate content - text, images, code. Not just analyzing inputs, but producing entirely new outputs. 6. Agentic AI Where things are heading. Beyond responses. These systems plan, take actions, and use tools with context. They don’t just assist. They execute. You can’t appreciate the top layer without understanding what supports it. We’re moving past the Generative phase into the Agentic era. AI is evolving into something that doesn’t just respond but operates. Those who master it will lead. P.S. Are you still learning prompts, or starting to build systems? ♻️ Repost to help more people understand what AI actually is. —— 📌 Get my top 100 infographics for free: 1) Follow me. 2) Subscribe to my free newsletter at WildCapital.co. You’ll receive them directly in your welcome email.

Anthropic CEO Dario Amodei on why everyone thinks Mythos was a PR stunt, and why the commercial reality tells a very different story: "We have suffered enormously commercially from not releasing this model. This model has incredibly accelerated research within Anthropic and production and next models. It would do the same in the outside world if we were to release it. This has hurt us enormously commercially." That is not the language of a company playing marketing games. Claude Mythos is Anthropic's most advanced frontier model. It has not been released to the public. And according to @DarioAmodei, that decision has come at a steep and very real financial cost. When asked whether Anthropic has had to make trade-offs they were not entirely comfortable with, his answer was direct: "Throughout the entire history of Anthropic has been trade-offs." In an ideal world, you would study every possible failure mode before releasing any model. Anthropic did delay the initial release of Claude. But only by a few months. Everything, he says, is a trade-off. What has shifted now is their position in the industry: "Now that we're in what I would describe as a commercially leading position, we can afford to move the dial even further towards being careful. That's what the Mythos release was about. It's very hard to do something like that if you're not the leading player." This is the part that gets overlooked in the PR stunt narrative. Holding back a model that could accelerate your revenue, your research pipeline, and your competitive position is not a move you make for optics. It is a move you make when you genuinely believe the stakes are high enough to absorb the loss. Anthropic absorbed the loss. Most companies treat safety as a checkbox on the way to shipping. Anthropic treated it as a reason not to ship at all, and paid for that conviction with real commercial damage.

Maggie Vo on what parents get wrong about teaching kids to use AI: Most parents think they need to protect their children from AI, or hand it to them and step back. Maggie suggests a third path: sit down with your kids and explore AI together. "I recommend sitting down with your kids and going through AI together. Ask a question and then say, 'This was really confidently said, but is that enough? When someone says something confidently, is that enough for you to believe that?'" The answer she's hoping for? No. Because the real skill is knowing what to do with one. She argues the most valuable thing a parent can model for their children isn’t the right answer, it’s the process of not knowing one. Show them what it looks like when an adult encounters something unfamiliar, and walk them through how you’d go about finding out. "What I want to impart upon my kids is that finding the answer is just the start of your learning journey." This runs counter to how most schools measure progress. Getting to the answer is what's tested. But Maggie sees that framing as the problem, not the solution. When the answer becomes a starting point rather than a finish line, something opens up. Especially when AI is involved. Every output becomes a prompt for the next question: Where else can I check this? What information do I actually need? Does this hold up? She's also clear about what she doesn't want: kids who simply trust things at face value. "Adults don't always have the answers. Kids are super smart and they're able to find answers on their own. If we just talk and ask the right questions, they'll be able to discern for themselves what is true and what isn't." The goal isn't to make children skeptical of AI. The goal is to make them thoughtful about any source of information, human or machine. AI just makes that lesson more urgent, and more teachable.

Warren Buffett on why holding too much cash is just as dangerous as holding too little: Buffett is asked about his view on holding cash as a strategy. He doesn't mince words: "Cash is always a bad investment." But he's quick to draw an important line. The problem isn't cash itself, it's excess cash: "When people said cash is king a year ago, I mean, that's crazy. Cash wasn't producing anything and it was sure to go down in value over time." That said, Buffett isn't arguing for zero liquidity. He frames it as a matter of survival: "You always want to be sure you have enough. It's like oxygen. You want to be sure it's around, you know, but you don't need to have excessive amounts of it around." Too little and you can't operate. Too much and you're quietly bleeding value every single day. He then shows how this plays out in his own decision-making: "We came in with something over 40 billion of cash and we've got about 20 billion now and we've had some earnings. So we put a lot of cash to work and I like that." Deploying over $20 billion wasn't a reluctant move, it was the goal. For @WarrenBuffett, sitting on surplus cash is simply a mistake: "Anytime we have surplus cash around, I'm unhappy." His preference is clear: "I'd much rather own a good business than have cash." Cash belongs in the hierarchy as a tool for survival and opportunity. The moment it becomes excess, it starts costing you.

Jensen Huang: "No matter what your dreams are, you're going to get punched in the face at some point."

What happens when consumer tech and automotive engineering collide at full speed? China's AI is showing us. Source: Douyin

In the movie Contact, aliens are asked one question: "How did you survive your technological adolescence without destroying yourselves?" Dario Amodei, quoted that scene because he thinks it's the question we now have to answer about ourselves.

Setting up is not starting. Planning is not building. And watching tutorials is not work: I spent months picking the "right" tools before I ever made a dollar. You know what finally worked? I picked one platform. Solved one problem. And asked one person to pay me. Most people are one boring decision away from a completely different life. 🎁Get PDFs of my top infographics here: fullpotentialzone.beehiiv.com/subscribe Please repost to help others!♻️

Brian Chesky on why you should never focus on a million users: Most founders get paralysed trying to build the next Apple or Google from day one. Brian Chesky's advice cuts through that: "Don't focus on a million. Focus just on a hundred." He explains that trying to scale to a million immediately forces you to build systems and processes before you're ready, and the complexity becomes unmanageable. So instead, you break it down: Get to 100. Then get to 1,000. Then keep going in orders of magnitude. "The job changes as you do that." The magic of this framework is that it makes the problem small and solvable at every stage. You're never trying to solve for a million. You're only ever solving for the next milestone. And when founders push back, pointing to Apple or Google as proof that you need to think big from the start, @bchesky has a ready answer: "Apple started by selling blue boxes out of the trunk of a car. Google was a research project they were going to sell for low millions of dollars and they didn't really know what they had." The companies that became giants didn't start as giants. "These things all start as unprestigious toys that seem hacked together and they're only made for you and your friends. That's almost always how it starts." The pressure to build something impressive from the beginning is the very thing that kills most startups before they get a chance to grow.

Perplexity AI CEO Aravind Srinivas on why every breakthrough in AI makes their business stronger without them building a single thing: Perplexity's revenue has more than tripled since the beginning of the year. But here's what's interesting about that growth — Perplexity didn't build the models that made it possible. Aravind explains: "Every time any part of the AI stack improves, our product improves." He breaks down exactly how this played out in practice: "Since the beginning of the year, Anthropic's models have made tremendous progress. But what's also true is that our revenue has more than tripled since the beginning of the year. A lot of thanks to model progress made by Anthropic and we also brought our burn down thanks to OpenAI competing with them and bringing down the cost of the same capability." And the compounding doesn't stop there. With open source and local models gaining ground, Perplexity is now looking at moving some inference back to local devices, cutting costs even further without lifting a finger on chip design. @AravSrinivas summarises the business logic behind this: "Every time any part of the AI stack, whether it's chips, models, harnesses, any of these, gets better, our system improves tremendously. And if our system improves tremendously, our users love it and they pay more. They spend more and so our business grows." This is the compounding flywheel that most people miss when they think about who wins in AI. The winner isn't always the one who builds the most or spends the most. Sometimes it's simply the one whose business benefits most from everyone else getting better. Aravind puts it plainly: "If Jensen produces a better chip, it's great for us. If Dario produces a better model, it's great for us. If Apple produces a better device, it's great for us. I love the fact that we are able to be a very positive player at every layer of the stack and not have to rely on any one person to win."

Demis Hassabis on why AI still hallucinates and the specific fix that's coming: Even Gemini 3, DeepMind's most capable model yet, still gives confident answers when it should admit uncertainty. Hannah Fry asked Demis whether you could build a system that produces a confidence score the way AlphaFold does for every protein structure it predicts. His answer: yes, and it's one of the missing pieces between today's models and AGI. "I think the better the models get, the more they know about what they know, if that makes sense. And I think the more reliable you could sort of rely on them to actually introspect in some way or do more thinking and actually realise for themselves that they're uncertain or there's uncertainty over this answer." The diagnosis of where hallucinations actually come from: "Sometimes it sort of forces itself to answer when it probably shouldn't. And then that can lead to a hallucination. So I think a lot of the hallucinations are of that type currently." Fry pointed out that under the hood, these models are already producing probabilities for the next token. Demis agreed but explained why that isn't enough: "That doesn't tell you the overall arching piece is, you know, how confident are you about this entire fact or this entire statement? And I think that's why you'll need this. I think we'll need to use the thinking steps and the planning steps to go back over what you just output." His analogy for what current models are doing wrong: "It's a little bit like the systems are just, it's like talking to a person and they just, you know, when they're in, on a bad day, they're just literally telling you the first thing that comes to their mind. Most of the time that'll be okay. But then sometimes when this very difficult thing, you'd want to stop, pause for a moment and maybe go over what you were about to say and adjust what you were about to say." The fix isn't a bigger model. It's getting the model to use its inference-time reasoning to audit its own draft before committing to it. The equivalent of a person catching themselves before they say something wrong. AlphaFold already proved a version of this is possible. Every structure it predicts comes with a per-residue confidence score that tells scientists which parts to trust. The challenge is doing the same thing for open-ended language, where the unit of "a fact" is much harder to define than the position of an atom.

Elon Musk drops by his team's livestream before Flight 11 and his excitement for the launch is contagious. Source: SpaceX

Scientists at Cortical Labs just grew 200,000 human neurons on a microchip and taught them to play Doom. Source: Cortical Labs

The smartest students in the AI era won't be the best at finding answers. They'll be the best at finding the right questions to ask.

"For me everything in my life was about becoming free." Pavel Durov on why he really built Telegram:

The 100x org went viral. Half the internet hated it. The other half was curious.   One month later: output is up. productivity is spiking. we're approaching a 5:1 agent-to-human ratio.   And contrary to popular belief, we're doing the OPPOSITE of tokenmaxxing. We're tokenSAVING....   Why 5,000 agents for 1,000 people? Because agents AREN'T workflows... Agents define permissions and skills. the vast majority of ours don't bother a single human, they run as triggers and loops in the background.   My rule is simple: if you're doing work that ISN'T purely using your judgment → build an agent or a system to remove that bottleneck.   So that's what we did. And yes, it broke things. Honestly it created real pain for first couple weeks. To be expected.   we have TONS of examples now people who had NEVER built a system in their life and are now running large-scale agentic systems that power much of our company.   I always say - if you automate your own job with AI, you'll ALWAYS have a job. Systems is arguably the most important thing of all.   Today, regardless of what your role used to be, everyone does the IC work end-to-end: → Before, you riffed on messaging. Now you ship the landing page yourself. → Before, you were a PM waiting on design and research. Now you're a builder iterating in real-time. → Before, you waited on someone to finish something. Now you don't.   The 100x future isn't coming. It's HERE. Bottlenecks are mostly gone and those that aren't are in sight.   And obviously gotta lead by example. I even designed and built the Brain² landing page. I built most of the product. The in-app announcements. The emails. All of it. Breaking our existing processes let us rethink everything WITHOUT prior constraints.   Enter, Process Miner: my agent that mines all the processes in our company. This is just one of the infinite use cases for setting up an LLM to process EVERY SINGLE event that happens in your company.   These types of agents simply are not sustainable without compressed context. Instead of starting from scratch each time and deep researching across the ~100,000 daily activity items in our company, it starts from context that has already been processed, summarized, and organized.   Everyone's jumping on the context and memory train now. I love it. We've been building this for two years. and we STARTED OVER 4 times! We came out the other side with what we call live intelligence: → LLM pipeline: every event runs through a cheap LLM that summarizes, organizes, triggers, and rolls up → Self-organizing memory. Don't overcomplicate it. Start simple. → Self-improving orchestration. THIS IS THE KEY. BUT... don't over-focus on orchestrating models... it's WAY more about orchestrating CONTEXT.   It wasn't more tokens. It wasn't a better model. It was CONTEXT.   Every LLM is average by default. Think about that. If you're using the same model as everyone else, your intelligence is by definition, AVERAGE.   That's what really changed everything for us internally and it's why we even have a product to release in the first place.   Our company AI is personalized (and yes these learnings are applicable regardless what system you use). Every time you use Brain², we dynamically rewrite the system prompt based on real-time context and intent. Your activity, your decisions, your projects, your memory, your preferences. And your FEEDBACK.. that's WAY more important than decisions that everyone goes overboard on being so valuable. All injected AHEAD of time.   When AI has 100% context, you don't need prompts in the same way. You need INTENT.   I go anywhere in ClickUp and type "@brain slides." Done. "@brain campaign." Shipped. "@brain app." Live.   And it REMEMBERS. Every interaction compounds. One Brain for your whole company, and every person makes it smarter for everyone else.   I flew our AI team to my house for a week to finish this. Monitors and whiteboards everywhere. SO proud of this 100x team.   Brain² is truly the best work AI in the world... and I firmly believe it's also the best personal AI. Naturally, easy for me to say and honestly I assumed it was just my bias originally. but then EVERYONE who tried it started saying the same thing.   So we tested it. A three-week study, randomly recruited participants, blind pairwise comparisons. They picked Brain² as the best nearly 100% of the time over ChatGPT, Gemini, and Claude.   Here's the kicker: you can run ANY of those models inside Brain². Same frontier models. Radically different results. Brain² doesn't win because the model is better. It wins because of CONTEXT. It learns from every single interaction and new piece of context it sees.   OH and we built our harness around being OBJECTIVE. Soo it can be sassy - you will definitely see some 'fun' answers in your day to day. AI that is built to agree with you is NOT productive. That makes no one better.   Run it against whatever you're using right now. It compounds in quality the more that you use it. Prove me wrong → clickup.com/brain Brain² is your company's AI. I built this for us originally to realize 100x, and now we're shipping it externally. Even if you're building your own I just hope sharing in public helps in some small way.   💜 Let's make the world more productive

Every bit of this article is pure gold. One of those pieces that I genuinely hope stays hidden:

Pascal - The X Growth Guy @xgrowthpascal

2 days ago

x.com/i/article/2053…

3 18 40 16K 82

Warner Music CEO Robert Kyncl shares why every musician should stop fearing AI. Source: Fortune Conversations (2026)

Boris Cherny put it perfectly: AI can answer questions only the top students in the world can solve then fail a logic problem a child could handle. That inconsistency is exactly what separates LLMs from true general intelligence. Media: Hanna Fry