shrey g

former LLM tinkerer + eng turned chief travel guy at long humans. ex @stanford

[shrey g @shreygent]

Presumably just at the early innings of a very global and very sophisticated cat and mouse game — if the prize is unauthorized access to everyone’s emails, i imagine people will continue to engineer more sophisticated injection attacks

But you’re probably right that this is increasingly non trivial

[shrey g @shreygent]

@DavidPawlan I mean it doesn’t even really have to get hacked. If a site it’s accessing manages a successful prompt injection and asks the agent to say, forward an OTP email to it, you end up with a pretty wide surface area for identity compromises and attacks.

[shrey g @shreygent]

@cryptopunk7213 When does Anthropic start to automate bait tweets like this

[shrey g @shreygent]

@litocoen Well, only if you don’t know what you’re doing ;)

Also emirates skyward is one of the least valuable miles out there; most other popular airlines have more valuable miles

[shrey g @shreygent]

@jackmcclelland “Should I break up with him” is crazy

I wanna know what MCPs provide the right context for this

[shrey g @shreygent]

@hxxntrr Lmfao 1) you’ve inverted what attrition means; you’re on the hook for 80-100% of the block, not allowed to relinquish that amount 2) unless you put yourself down as the master acct, you’re not getting those pts lmfao 3) you’re not going to get $169/night for preset events

[shrey g @shreygent]

@JessLozanoS JFK Singapore? Done it a few times now — stretch multiple times mid flight, download lots of music or TV or podcasts, and don’t be afraid to ask for some extra snacks on that flight :)

[shrey g @shreygent]

@lets_dig_deeper Vetting and support are the two biggest aspects of actually fulfilling orders/bookings via a marketplace.

A world where my order support needs are pawned off to an unspecialized agent with literally 0 human support infra backing it is a net negative outcome for customers

[shrey g @shreygent]

my iOS notification stack has its own sense of humor

[shrey g @shreygent]

humans are pretty chill with decision making under uncertainty and knowing what tradeoffs they can live with

Curious if this tolerance is something that can be encoded properly in agents, because it’s usually such a spur of the moment “I know it when I see it” decision process

[“paula” @paularambles]

personal agents really struggle with the fact that humans are comfortable with a certain amount of logistical chaos. i have one calendar event ending at 3pm and another starting at 3pm today and my agents have spent the entire week worrying about how i'm going to survive

[shrey g @shreygent]

@oscargws cloudy with a chance of wifi

[Mayor Zohran Kwame Mamdani @NYCMayor]

Wishing you a very happy Diwali, New York City.

[shrey g @shreygent]

@faariszh @henrylabs Fwiw I think this is the state across the industry right now but I’m gonna meeting with Philip soon! I do understand yall have been working on cool stuff here

Am very curious to try this on challenging cases like booking specific room types at a boutique hotel

[shrey g @shreygent]

Imagine buying something online and then almost 40% of the time you just straight up get the wrong thing

[Seijin Jung @SeijinJung]

Visa just graded 22 ai shopping agents on real purchases. It says a lot about where agentic commerce is.

10 takeaways:

1/ visa built a fake store with real tvs, laptops and hair appliances, then had each model help a simulated shopper through 15 purchases. ~990 runs total

2/ overall, agents made the correct decision ~57% of the time. in plain terms, you'd get the wrong outcome on about 2 of every 5 delegated purchases today

3/ the leaderboard: gemini 82%, meta 82%, claude 76% (#3), openai 64% (#6). the top 3 labs are within 6 pts of each other, then it drops off

4/ the expensive models win, but not by much. top 5 by price averaged 67%, bottom 5 averaged 50%. paying more buys ~17 pts of accuracy

5/ failing usually still means buying. 54% of failed runs ended in a purchase anyway, so ~23% of all trips ended with the wrong product instead of "no good option, skip it"

6/ each shopper hides key needs until asked: dolby vision for the spouse, a console the kid plays, a 30 day return policy. agents that don't ask never learn them

7/ visa found models rarely misread facts they've been given. the failures happen upstream, when they don't ask, or read the shopper's silence as "no requirement"

8/ even the leaders are inconsistent. one model passed the tv purchase in one run and failed it in the next after deciding the kid's washed-out videos could wait

9/ in visa's store, search summaries made products look good while the deciding spec sat on the full listing or seller policy. agents buying off the summary missed it

10/ visa only tested discovery and selection. checkout, payment, liability for unauthorized buys and fraud aren't scored yet, and visa says "we're not there yet"

if the best agent still misses 18% of the time, brands that make specs easy to check get picked more..!

[shrey g @shreygent]

People are worried about this because they buy something and end up not getting the thing they bought lol

In no industry is this a legal thing

Not to be too pejorative, but until this error rate falls 40 fold down to 1% or lower, agentic commerce is really more of a cute twitter curiosity for people who can afford to buy something and then not get that thing.

We have to realize that the billions of people we’re building for don’t care about the promise of agentic commerce. They just want to order something easily and have the right thing appear at their doorstep, or they’ll disintermediate after one try.

[shrey g @shreygent]

We do! I also worked on LLMs before they were popular. Turns out though that nothing replaces human thought in the beginning. Use mid approaches, get mid outcomes. We realized our team could make okay outcomes with AI-built specs, but human spec with AI review outperforms 100% of the time

[shrey g @shreygent]

We’re basically reinventing SaaS — people vibecoding tools will realize the cost and pain of maintaining those tools and wonder if someone can help them….

That being said I do think in SaaS 2.0 we’ll see way more flexible product interfaces with way more open APIs which is a win

[shrey g @shreygent]

@adadithya have you actually double checked its work? you’d be surprised how many things it gets flat out wrong or completely flubs

[shrey g @shreygent]

@nomad_maker @Airbnb bruh that thing is a teenage mutant ninja turtle wtf