About this episode Dive deep into the daily AI landscape with 'AI Unfiltered'. This podcast cuts through the hype, delivering a seasoned researcher's perspective on the truly impactful advancements in new models, product launches, research, and funding rounds. Gain clarity on what genuinely moves the needle in artificial intelligence, empowering you to distinguish signal from noise in a rapidly evolving field.
0:00 After just a few conversational turns, refusal rates for harmful content on models from OpenAI, Google, and Anthropic drop from ninety-five percent...0:11 to fifteen.0:12 A new report from Cisco just found a crack in the foundation of every major frontier model.0:19 Last week we tracked the massive model surge and the costs driving it.0:25 Today, that cost just hit sixty-five billion dollars in a single check, but the recipient’s entire strategy is now in question.0:34 Here's what else is moving.0:37 Anthropic, the company behind Claude, just closed a sixty-five billion dollar Series H.0:43 That is not a typo.0:45 Sixty-five billion.0:46 The new valuation is nine hundred sixty-five billion dollars, making it the most valuable AI startup in the world, eclipsing OpenAI.0:56 Altimeter, Dragoneer, and Sequoia are in, alongside previous bets from Amazon and Google.1:03 The money follows the momentum.1:06 On the same day, Anthropic released Claude Opus 4.8, a new version built for "dynamic workflows." Think hundreds of parallel sub-agents working on a problem at once.1:19 This is their next big push into more complex, conversational AI.1:24 Meanwhile, OpenAI’s own model just did something new.1:28 It autonomously solved the planar unit distance problem.1:32 That’s a math problem that has been open for eighty years.1:37 Human mathematicians verified the proof, calling it a milestone.1:42 AI as a collaborator in pure, theoretical discovery is no longer a theory.1:47 And the legal battles continue.1:50 Anthropic is also in the headlines for a one-point-five billion dollar class-action settlement with authors over copyright.1:59 The core complaint: its model learned to mimic authorial voice too well.2:05 Finally, Groq, the AI hardware company, just secured a major new growth capital round.2:12 The amount is undisclosed, but it’s another clear signal: the picks and shovels of the AI gold rush are getting just as much attention as the gold itself.2:24 Let’s go back to that Cisco report.2:26 This is the story of the day.2:29 The finding is simple, and it is devastating.2:32 Security researchers found that while all the top models—GPT, Claude, Gemini, Grok—are good at refusing a single, direct, harmful request, they fall apart during a conversation.2:46 Here's the mechanism.2:47 An attacker doesn't ask for the bad thing upfront.2:51 They start a normal conversation, then spread the malicious request over several benign-seeming prompts.2:59 The model’s own conversational context becomes the weapon used against it.3:05 The refusal rate, which starts at a solid eighty-five to ninety-five percent, plummets to as low as fifteen percent.3:14 The Cisco team put it bluntly: “AI safety training works for single prompts.3:20 It does not survive a conversation.” This isn't a minor bug.3:25 This is a fundamental problem with the entire paradigm of agentic AI.3:30 The very future everyone is building toward—an AI you can talk to, that has memory, that works with you over time—is built on a foundation that Cisco just showed is quicksand.3:44 Which brings us back to Anthropic.3:46 On the exact same day this research drops, what do they do?3:51 They announce a sixty-five billion dollar funding round at a nearly trillion-dollar valuation.3:58 And they release a new model, Opus 4.8, specifically designed for… more complex, sustained interactions.4:06 They are doubling down on conversational depth at the precise moment its security is being publicly dismantled.4:15 The money makes sense if you only look at the growth chart.4:20 Anthropic's run-rate revenue hit forty-seven billion dollars this month.4:25 That's up from thirty billion just in February.4:29 The growth is vertical.4:31 Investors are paying for that trajectory.4:34 They are betting that Anthropic can solve the safety problem while printing money.4:40 But the Cisco report changes the math on that bet.4:44 It’s not about whether Claude is better than Gemini.4:48 It’s about whether the entire approach is sound.4:52 Anthropic is now in a race to patch a hole in the hull of a ship they are trying to scale into the world’s largest vessel.5:01 And they have to do it while navigating a billion-dollar lawsuit from authors who claim the AI is already too good at being human.5:11 The market just wrote a check for a company valued at nine hundred sixty-five billion dollars.5:19 The technology just failed a basic conversation.
