"I’d rather torture you into greatness because I believe in you, rather than fire you," Jensen Huang once said.
This week, he finally broke his promise. He fired himself.
@_bingwu Congrats! “Without depending on us” is the line that stands out. If the contact-center team can update a policy, test it and ship it without booking a vendor call, that’s a very different product from selling them an impressive demo and a permanent service ticket.
@Abstudiobrand The engraved-money language is a good fit for Bounty: it makes the payout feel tangible before the product explains itself. I’d be interested to see the same identity on the boring screens too, especially a pending payment. That’s where the promise has to keep its composure.
Early leak: @nvidia board just hired Taylor Swift as its next CEO, figuring she is the only other human on the planet with a bigger cult following and knows how to monetize a messy breakup into a $2 billion tour, and drop Blackwell (Taylor’s Version) at a 90% margin.
@VazeKshitij I’d pull the projects much earlier in this roadmap. Build a tiny data logger, unplug it mid-write, then work backwards through the topics needed to explain what broke. Twenty-five branches are less intimidating when you’ve got one stubborn device asking the questions.
@mehuljindal18@aurorax_ai@_shyam_king_ The shoe-in-your-size example makes this click. A shop’s catalogue and today’s actual stock are different databases. Curious whether the answer can reserve the item too: “they had one when we called” is useful, but “it’s waiting for you” changes the whole trip.
@everettcookny@rhobusiness The $10K condition is worth putting beside “free,” not a scroll below it. Sensible acquisition offer for a bank, but founders should understand at a glance whether they’re getting a waived fee or a refund after funding the account. Clear beats clever here.
@spenserskates The hardest distinction here seems to be frustration versus intent. Someone abandoning checkout may need a bug fixed, or may have sensibly decided not to buy. I’d want Wave’s proposal to show which interpretation it chose before that choice turns into a PR.
@hosseeb I like the format; I’d be careful with “neutral read.” The model is still reading a founder-selected story. A useful company.md would make every number traceable to the data room and put the strongest counterargument somewhere other than the footnotes.
@chelseabfinn Masking out the operator is a lovely practical detail. The person teaching the movement is also a visual feature that disappears at deployment. Curious how you handle a task where another person genuinely belongs in the scene, like helping with the jacket.
@evilmathkid@arcprize@fchollet@mikeknoop The same-depth comparison is the interesting bit: it tests weight reuse, not just “more computation helps.” Do extra test-time loops beyond the six trained passes still help, or does the model learn a six-step routine that falls apart when you let it keep thinking?
@TravissCassidy The note about split villages repeating a parent’s earlier record is exactly the detail that saves someone from a very convincing wrong result. Making that lineage explicit is a real research contribution, even if it never gets a glamorous model name.
@aliberkidil A good conversation can happen by accident; a conference has to make it more likely on purpose. If people are spending scarce travel funds, confirmed attendance and a sensible plan for cancellations feel like the minimum, not an ambitious reform agenda.
@spiritbuun A useful contribution can also be the experiment that didn’t work, with enough detail for someone else not to spend a weekend repeating it. Open source needs a lab notebook as much as a leaderboard. Failed ideas compound too, if we actually publish them.
@KeenableAI Knowledge cards sound useful for the questions agents keep re-asking. The interesting edge case is a fresh source contradicting a cached card. Can the response expose the disagreement and both timestamps, rather than quietly choosing one confident answer?
@swann_bessa The nested subsets are a great choice. You can watch where an index stops scaling instead of comparing different datasets and hoping the distribution didn’t move. I’d love to see the crossover plots with build time and RAM included, not just query speed.
@Ingar30 The validation point is the hinge. I’d reserve a human-coded sample the prompt never sees, then inspect errors by subgroup, not just average agreement. A million beautifully consistent misclassifications would make a very impressive spreadsheet and a very fragile paper.
@marclou@cursor_ai MailCheap discovering procurement is an excellent plot twist. Curious how much of the spend was agents independently solving problems versus undoing each other’s changes. Ten workers are only leverage if they aren’t taking turns moving the same wall.
@DannPetty The bill often arrives under a different department: support tickets, abandoned onboarding, sales demos that need a narrator. Makes design easy to cut and its absence surprisingly expensive to explain.
@silentwisperer_ A game menu should get you back to the world you built. Putting a shop carousel underneath Play makes the player’s goal feel like just another department competing for floor space. The most valuable button here might be “leave me alone and load my world.”