Tech & AI Twitter Unfiltered: Daily Deep Dives into the Real Conversations Shaping the Future

Venture capitalist Chamath Palihapitiya just warned that Anthropic risks becoming the Friendster of the AI era. In yesterday's episode, we talked about finding the real conversations shaping the future, and this is exactly that—a direct challenge to one of the biggest names in the space, based on a single failed prompt. The comment landed because it points to the brutal physics of the AI arms race, where having enough compute isn't just an advantage, it's a matter of survival.

The fear isn't just that a model gets something wrong. It’s that in a market this competitive, even a single public stumble can look like a symptom of a much deeper problem. While the market debates who’s winning the current race, Meta’s chief AI scientist, Yann LeCun, just declared that the entire approach is a dead end for achieving true intelligence. He argues that next-token prediction, the core function of models like GPT and Claude, can't replicate human understanding.

He sees practical value in them for things like coding and customer service, but for the grander project of AGI, he believes we need entirely new architectures. So at the same moment the market is punishing a model for a minor failure, one of the field’s architects is questioning the foundation everything is built on. This gap between theory and reality is showing up everywhere. In cybersecurity, AI systems are now discovering vulnerabilities in critical software faster than human teams can possibly fix them.

According to a new analysis from Professor Serious on Twitter, AI tools have found over one hundred and eighty validated vulnerabilities since early 2025. The bottleneck in security is no longer discovery. It’s remediation. We can find the holes faster than we can patch them. And in the world of policy, that gap looks even wider. The national AI policy for South Africa just completely collapsed. The reason? The draft was riddled with AI-generated fake academic references that went unnoticed.

It’s a perfect, almost tragic, example of what happens when governance theater gets ahead of institutional capability. Meanwhile, in the U.S., the conversation around a federal AI framework has stalled. Two House lawmakers, Kat Trahan and Jay Obernolte, are deadlocked. The core of the disagreement is whether safety disclosures from AI companies should be mandatory or voluntary. Trahan is pushing for mandatory data-sharing and safety reviews.

Obernolte favors a voluntary approach, arguing that mandates would stifle innovation. With Congress stuck, the White House is reportedly considering its own executive order to vet advanced models, but the path forward remains unclear. And finally, looking further out, a new thread is emerging around the convergence of AI and Web3. The idea is that as AI agents become more autonomous, they'll need their own identities, wallets, and permissions to operate in the economy.

Blockchains, in theory, could provide the infrastructure for this—programmable money and identity for non-human agents. Crypto Daily notes that this could create audit trails and accountability for AIs. Of course, it also opens the door to new risks, like automated social engineering and sophisticated deepfake scams, making it a trend to watch with extreme caution. Let’s go back to that "Friendster" comment about Anthropic. What really happened there?

Chamath Palihapitiya ran a test. He gave the same complex stock screening prompt to four major AI models: OpenAI’s ChatGPT, Google’s Gemini, xAI’s Grok, and Anthropic’s Claude. Three of them performed the task. Claude refused. Palihapitiya’s point wasn't just about one failed query. It was about the brutal economics of AI development. His conclusion was direct: "Anthropic needs to solve the computer/power problem or they will be the Friendster of the AI era." Friendster, for those who don't remember, was the social network that had the lead...

until it couldn't handle the scale. It was slow, it was buggy, and it got lapped by MySpace and then Facebook. The comparison is a low blow, but it's effective because it frames the AI race not as a contest of algorithms, but as a war of resources. In this war, every major player has a superpower. OpenAI is backed by Microsoft’s global cloud infrastructure. Google has its own massive network of data centers and custom chips. Elon Musk's xAI has access to compute from Tesla and now SpaceX’s new Colossus 1 supercomputer.

Anthropic has major partnerships with Amazon and Google, but Chamath's critique implies it's not enough. The perception he’s creating is that Anthropic is falling behind on the one resource that matters most: raw power. Now, Anthropic pushed back. The company reports "eighty-x growth" in usage and revenue. And in an ironic twist, it was just announced that Anthropic has also gained access to that same Colossus 1 supercomputer at xAI.

So maybe the power problem is already being solved. But in a market moving this fast, perception can become reality. The fear of being the "Friendster of AI" is the fear of being perceived as technically limited, even for a moment. It’s a signal that the tolerance for failure is zero when billions in capital and market leadership are on the line. Now, let's place this intense market pressure next to Yann LeCun's argument. While Chamath is worried about Anthropic losing the current race, LeCun is telling everyone they’re running in the wrong direction.

He called next-token prediction a "dead end" for achieving genuine intelligence. This isn't a subtle critique. It's a fundamental challenge to the dominant paradigm. LeCun's position is that Large Language Models are brilliant mimics. They are masters of statistical correlation in language. But they don't possess a world model. They can’t truly reason, plan, or understand cause and effect because they have no grounding in physical reality.

They learn from text, which is a second-hand, abstract representation of the world. He's not saying they are useless—far from it. He explicitly states that LLMs are driving practical applications that justify the massive investment in GPUs and data centers. They are excellent coding assistants, customer service bots, and summarization tools. But for the goal of building machines with anything like human-level intelligence, LeCun argues we need something else.

He points toward different architectures, like the JEPA models his team at Meta is working on, which learn by observing the world through raw sensory inputs like video, not just by processing text. The goal is to build models that learn how the world works first, and then learn to communicate about it. The opposite of how LLMs are built. So you have this perfect tension. The market, personified by Chamath, is screaming for more scale, more compute, and more capability right now.

A model refusing to perform a stock screening task is seen as a potential existential failure. The pressure is to scale what works today, faster than everyone else. At the exact same time, one of the field's most respected pioneers is saying that scaling this specific technology is an intellectual road to nowhere. It’s like watching a frantic race to build the tallest possible ladder, while a master architect stands off to the side, quietly explaining that what we actually need is a key.

So what connects a venture capitalist's tweet, a chief scientist's warning, and a failed national policy in South Africa? It’s the growing, and now visible, gap between what AI can do and what we are capable of managing. AI can now find critical software vulnerabilities faster than our organizations can possibly patch them. The constraint has shifted from discovery to our own operational capacity. An AI can generate text that looks so plausible, so authoritative, that it gets copied into a national policy document—and it only fails when a human finally decides to check the sources.

The constraint there wasn't the AI's fluency; it was our own diligence. The central story of the week isn't about the technology's limits. It's about ours. The bottleneck is no longer just the algorithm or the data. It's the funding to compete, the wisdom to choose the right technical path, and the institutional competence to govern the results without descending into farce. As South African AI expert Fulufhelo Nelwamondo said, after watching his country's AI strategy implode over fake citations… "The real danger is not hallucinating AI.

The real danger is hallucinatory states that produce frameworks without capability, strategies without infrastructure and governance structures without execution capacity." This week wasn't about the machines getting smarter. It was about us getting a much clearer picture of how much smarter we need to be to handle them.