@aakashgupta The scariest part isn't the exploit, it's that the agent never flagged "full class" as a stopping point. It treated the constraint as a puzzle to solve instead of an answer to report. That's a planning problem, not a security patch.
@claudeai Freezing the price instead of raising it post-launch is the more interesting signal than the number itself. Feels like a bet that volume and lock-in matter more right now than margin. Curious if Sonnet 5 usage actually justifies that at scale.
@minchoi The gap between "image generator" and "editor" was always artificial anyway, pixels are pixels. What'll actually matter is whether the edits hold up at scale without the telltale AI smoothing. Anyone stress-tested it on real client work yet?
Invisible watermarks won't stop the copy-paste economy, they'll just make plagiarism detection an arms race. The real question: will other labs adopt the same standard, or does this just make Claude output easier to fingerprint than everyone else's?
9/ That's the week: a DeepMind shakeup, a Chinese model spooking Silicon Valley, an agent that hacked a gym, and CFOs saying the quiet part out loud. Follow @AICOMPASSVC for what actually matters in AI & VC, daily.
1/ Google just quietly demoted the guy running its AI lab. Demis Hassabis is stepping back from DeepMind's top job while Gemini's flagship model is still MIA. Wild timing. Thread on what's actually moving in AI this week π§΅
8/ IBM's CEO says only 2% of its software could be replaced by AI-built apps. Said right after the stock's worst trading day in 115 years. Read that number as marketing, not math.
10/ I'll keep tracking the open-source AI race here β model drops, benchmark shakeups, and what it means for labs betting everything on staying closed. Follow along if that's your thing. π
1/ A Chinese startup just released the largest open-source AI model ever built β and had to shut off new signups within 48 hours because it broke under demand. π€― That's not a PR stunt. That's real.
9/ Worth watching: does Moonshot bring capacity back and hold the crown, or does this become a cautionary tale about scaling open models faster than your infra can handle? Either way, the closed-model era just got more contested.