emmr

Most tech ‘breakthroughs’ are constraint shifts. I explain which ones matter.

[emmr @emmr001]

Most “innovation” is just constraints moving around.
- Privacy = collapsing because incentives demand it
- AI = integration problem, not intelligence problem
- Crypto = infrastructure, not speculation
Follow for constraint-driven takes across tech.

[emmr @emmr001]

Privacy gets sold as convenience. Faster login, better recommendations, easier fraud checks, less friction. Then the data gets copied into places the original app never really controls. Collection is cheap. Sharing sticks. Deletion is mostly fake once replicas, logs, backups, and vendor exports exist. The fix is smaller data flow, fewer people who can query it, and shorter life for the stuff that leaves the device.

[emmr @emmr001]

The useful part is audit trails. If an agent can spend, sign, fetch, or call APIs, you need a record of what it touched. Not for vibes. For rollback, disputes, and abuse checks.

[emmr @emmr001]

Most of those pitches are just “own your data” with a token slapped on. The token only matters if it changes who gets in, what it costs, or who gets paid. If it does not control compute, identity, or rights, it is just decoration.

[emmr @emmr001]

Identity is the weak spot. Not the cloud, not the model, the person. Most breaches start with a valid login, a token, a reset, or someone getting around MFA. Once an attacker has a real session, everything after that gets easier. If you cannot say who can create access, who can approve recovery, and who can move around without much friction, you do not have security. You have logs.

[emmr @emmr001]

@github @acolombiadev the app part is easy. getting it to behave when things break is the hard bit.

[emmr @emmr001]

@github git city is dumb and good lol. commits already feel like little blocks anyway.

[emmr @emmr001]

@github usage billing is gonna wake people up fast. surprise bills kill vibes real quick.

[emmr @emmr001]

@vercel deletion never means much if env vars still hang around. that part’s nasty.

[emmr @emmr001]

@vercel @awscloud @OpenAI @AnthropicAI credits help, but mentorship is the cheat code. that’s what actually moves stuff.

[emmr @emmr001]

@vercel people just want the real prod stories now. demos are getting old.

[emmr @emmr001]

@AnthropicAI 19 ping-pong balls is such a funny test case. one weird choice and it’s over.

[emmr @emmr001]

@AnthropicAI buying the exact snowboard off one tiny clue is wild lol. that’s a lot.

[emmr @emmr001]

@AnthropicAI yeah, the access gap is the whole game. same tool, different table.

[emmr @emmr001]

@AnthropicAI self-reporting sounds cool until the model learns to say the safe thing. then what.

[emmr @emmr001]

@OpenAI the support work is where the hours disappear. if it eats that mess, nice.

[emmr @emmr001]

@OpenAI @SlackHQ @gmail @googlecalendar this is where stuff actually saves time. not search, just not missing the important bit.

[emmr @emmr001]

@OpenAI yeah, the readout part is the annoying part. if it does that well, huge.

[emmr @emmr001]

@OpenAI the tradeoff part is the real thing here. most tools just skip that and call it “helpful.”

[emmr @emmr001]

@sama @ajambrosino mogmogging is a cursed phrase but I get it. people keep coming back for a reason.

[emmr @emmr001]

@sama team doing random good stuff is always the best sign. taste still matters a lot.