So guess what: AI has now, OFFICIALLY, helped design a most cancers remedy that simply cleared Phase 3.
Here’s what’s up: Moderna says its AI algorithms take sequencing information from a affected person’s tumor and blood, assessment the most cancers’s mutations, and predict as much as 34 “neoantigens” most certainly to set off an immune response.
Those AI-selected targets are encoded right into a customized mRNA remedy made for that affected person, then paired with KEYTRUDA. In the melanoma trial, the combo considerably improved recurrence-free and distant-metastasis-free survival versus KEYTRUDA alone; overall-survival follow-up remains to be ongoing.
This issues loads to me personally, as a result of my cousin died from Melanoma. So ya, I’m gonna take the win.
Translation: are customized GPTs nonetheless a part of your day by day stack, or have they change into the AI equal of an outdated Chrome extension you swear you continue to use?
Here’s what occurred in AI at this time:
-
😺 Generalist realized a job from one 3-second demo.
-
📰 Anthropic handed OpenAI in quarterly income.
-
🎓 Cursor can auto-fix new PR suggestions.
-
🍪 Google gave college students a free 12 months of Gemini.
-
📰 Flock constructed AI to go looking police surveillance information.
Hey! We’ve solely received a couple of advert slots left in Q3 and they’re going FAST!
If you need to get your advert in entrance of 700Ok+ day by day readers who’re LOCKED in on AI and what actually issues, ensure you attain out ASAP w/ the button under.
P.S: need to get forward of the gang? Email Garrett immediately and he’ll get you arrange.
So y’understand how you and I can just about watch anyone do one thing as soon as, and no less than try it, even when we kinda suck at it? Well, now robots can do this too!
Yesterday, Rich Sutton argued that AI’s subsequent leap received’t come from stuffing fashions with extra human-made information. It’ll come from brokers that continue to learn from the world they’re truly working in. Less than 24 hours later, Generalist dropped GEN-1.5, a robotic mannequin that makes that concept really feel loads much less theoretical.
Here’s what occurred: GEN-1.5 can watch a single 3–12-second bodily demonstration and instantly try the brand new job. Generalist calls it “physical prompting”: the demo sits within the robotic’s 30-second context window, with zero gradient updates.
-
Across 10 easy duties, one demo produced 59% common success; 10 weight updates on 5 minutes of knowledge raised that to 83%.
-
It can copy some human-hand demonstrations, use simulated demonstrations on an actual robotic, mix two bodily prompts, and improvise with unseen instruments.
-
Previous robotic adaptation can take tens of 1000’s of gradient steps. Generalist says 10 steps modified GEN-1.5’s weights by lower than 0.15%.
Our take: This issues for 2 causes. First, the interface adjustments from programming a robotic to exhibiting it. Second, it begins to make Sutton’s “Big World” thesis concrete: the world is way greater than any static coaching set, so helpful intelligence has to continue to learn from no matter it encounters now.
There’s an vital distinction Sutton would care about: GEN-1.5’s one-shot trick is in-context studying, not persistent studying. Its weights don’t change. The few-shot mode, the place GEN-1.5 updates its weights in 1–10 steps, is definitely nearer to Sutton’s continual-learning imaginative and prescient.
The milestone isn’t 59%. It’s that eight months of broad bodily pretraining made a couple of seconds of latest expertise helpful. Generalist describes adaptation now as nearer to reminding the mannequin of one thing it almost is aware of. If robots could make these new expertise persist and compound with out forgetting outdated ones, Sutton’s “learn from the current world” future begins wanting loads much less summary.
Cursor’s new Subscriptions let cloud brokers monitor a…
