AI News · The field notes
The Daily AI Drop - Morning Edition, October 25, 2026
Alex and Jordan cover five overnight drops: Dessa, the speech model that learned to talk by listening, Sam Altman on what people still bring to an AI world, Nate B. Jones on AI agents as messy roommates, the new shape of AI coding interviews, and a web vulnerability affecting seventy percent of sites.
Alex and Jordan cover five overnight drops: Dessa, the speech model that learned to talk by listening, Sam Altman on what people still bring to an AI world, Nate B. Jones on AI agents as messy roommates, the new shape of AI coding interviews, and a web vulnerability affecting seventy percent of sites.
🎧 Listen: Listen to this episode
Fahd Mirza — Dessa by Jack Dorsey's Block
Fahd reviews Dessa, a speech model from Jack Dorsey's Block that learned to talk by listening rather than from scripted responses. That approach could make voice assistants feel far more natural, picking up rhythm, timing, and tone instead of just words. Voice is becoming the front door to AI, so whoever gets natural conversation right earns a lot of trust — the open question is whether Dessa holds up beyond the first minute.
Matthew Berman — Sam Altman on What People Still Bring to an AI World
Berman looks at Altman's argument about the human contribution in an AI world: it does not disappear, it moves toward judgment, taste, purpose, and responsibility for the outcome. When AI makes producing things cheap, choosing well becomes the valuable skill. If you can generate a hundred ideas in an afternoon, the person who can spot the good one — and answer for the choice — becomes more important, not less.
Nate B. Jones — AI agents = messy roommates
If you run agents all day, they are probably using your computer more than you are: opening browsers, starting services, writing files, and leaving it all behind when the job ends. A faster chip and more memory help, but they will not stop that load from growing. Nate's advice is to set house rules for the agents already on your machine — clean up after jobs and be deliberate about what an agent may start and keep — before spending money on bigger hardware, and he shares the memory and storage he would actually buy for an agent computer.
Aish Reganti — Still grinding LeetCode? What AI coding interviews test
Some of Silicon Valley's fastest growing startups have moved away from LeetCode-style interviews, checking four things instead: architecture, implementation choices, edge cases, and verification. That maps far better to real software work with AI in the loop than abstract puzzles under pressure. For job seekers, the new playbook is to practice explaining your design, showing your tradeoffs, testing weird inputs, and proving your code works — because in an AI-assisted world, checking the work is the work.
Priyanka Vergadia — 70% of the Web is Vulnerable Right Now
Seventy percent of the world's websites are vulnerable, headlined by the PHP2Shell flaw: one request, no password or login, and a hacker takes over a site. The fix is on the server, not in WordPress, so patching your own site alone may not be enough. Vergadia's practical advice is to turn on two-factor authentication everywhere it matters and watch for fake reset emails, since attackers love to follow a big vulnerability story with phishing that looks like help.