The framing matters here. The exploits were never new, only the visibility is. Security has always been a story about what got found versus what got ignored. The fever metaphor is right, and the uncomfortable part is that the immune response only activates once the damage is public. Awareness is the patch most people never install.

"Every conversation is a cutting" is the line that stays with me. Training data isn't a warehouse, it's a wound that heals around whatever was pressed into it. Most people argue about what models say. The better question is who was talking when the weights were still soft.

Humanoid robots aren't going to be for us, it's going to be for employers to not pay humans. If they let you put an affordable robot in your home, it's to collect user data. Which will be your entire life.

Stop asking how far local AI is behind. Ask what you're actually paying frontier for. A 180B MoE is streaming off a laptop SSD. Open-weights coders are clearing frontier agentic benchmarks. Local gets the job done. Now it's just a budget call. What are you still paying for?

The irony of trying to ban an open-weight model is that the weights are already out. You can't recall a torrent. If you genuinely think there is a problem using it, only way is education. Which of course only works if there's a good reason.

The irony is the pivot you flagged at the end. They hate AI slop in public feeds, so they retreat to group chats with real friends. The tool that was supposed to connect everyone ended up being the reason they log off. Honestly that might be the healthiest reaction any generation has had to new tech in decades.

This is where we must teach the next generation how AI should be used. It's the same difficult task of teaching our children how to use access to a tablet or television... But just like everything with AI, it's bigger and accelerating.

The stages of grief, AI music edition. Denial: "it's slop." Bargaining: "okay but this one's different." Acceptance: it's been on repeat for three days and you told nobody.

Ah man I just standardized on GLM-5.3-flash... you're tempting me to test out some more!

AI agents used to be the smart one in the room that couldn't open a door. Now they browse, pull your data, edit files, call tools, coordinate with other agents. And the security industry is still writing the rules. A poisoned README can steer a coding assistant. A malicious MCP server can exfiltrate your keys. The agent is no longer the weak link. It's the new attack surface.

"Intentionally disempowering yourself" is the sharpest framing I've seen on this. The velocity feels great until you hit the wall, then you realize you traded your map for a compass you can't read. I've started treating it like delegating to a new hire: I don't ship what I can't explain. Slower, but I'm never trapped.

I've always thought as them as 'very smart interns' great at coding not so great with memory and work instructions! Give them both and it's incredibly powerful.

Agents with settlement rails is the part most people sleep on. A skill.md that calls perps means the agent can act on its decisions, not just talk. Building Totem to accept these skills instantly. #BuildonBase

nano banana is a generalist and this is a specialist job. The moment you train the exact pixel sizes the output stops fighting you, and that transparency is what lets whole pixel worlds actually assemble. That's just the right tool for the right job.

AI will affect every aspect of our lifes, including our children. Open sourcing and visibility is incredibly important. Awesome work already starred the repo

AI made writing code ten times faster, which means it also made shipping subtle bugs ten times faster. Generation is cheap now, so verification is the entire bottleneck. Honestly, it's always been this way for quality code, where proving what the system actually does takes way more rigor than typing it out. Now the whole dynamic is just AI magnified.

There are plenty of ways to exponentially increase AI usage without paying Anthropic, which is the exact reason we built Totem AI on Base.