Elyas Masrour

Last week a respected VC posted a dangerous global conspiracy he found with ChatGPT It was entirely a hallucination, and the latest example of “ChatGPT‑induced psychosis” When I wrote this script, this problem was hypothetical. Now, it’s here. Here’s my new film: hallucinations

also @kthai1618

@max_spero_ irl

the only dosa maxxing ai lab

@zagdotdev this is sickkkk 🔥🔥🔥

Introducing Zag AI review agents for iOS, iPadOS, and macOS developers. Zag runs agents in real Apple developer environments with Apple Silicon, Xcode, simulators, and your complete toolchain. Describe agents in TypeScript or Swift that do code review, QA, security, or App Store compliance, and run them on every PR.

At this point, Pangram's generalization is no secret. We train our model on millions of documents written by AI and humans, and the detector learns the underlying difference. Training on new models can help performance, but we often see very good accuracy right out the gate!

Opus 4.7 just dropped. Pangram hasn't trained on it. Does Pangram catch it? Yes!

@keysmashbandit A little bit but not that much!

@elyasbuilds Is the diversity of emojis that LLMs use smaller?

What's your favorite emoji? Do you use 🙄 or 🤣? You're probably a human! Prefer ✨ or 🚀? We have bad news...🤖 Pangram has millions of human and AI documents in our training set. Here, we analyze which emojis appear in different types of text!

Emojis aren't the key to AI detection. But tracing the patterns between AI and human text is a necessary part of the work we do here at Pangram. We're hard at work on new updates to give our users more clarity into the origins of their content. On that front, stay tuned! 👀

On the other hand, humans use emojis of faces 7.1x more than AI does 😌😛😚 Isn't that sweet?

@sadjamz_ @Google nice take. however. aws sucks so bad. and still it wins every day of the week.

@JustenMichel self actualization is important!!

it's interesting that people prefer guaranteed jobs over guaranteed income.

@taoburr Seems like a really well thought-out policy. Responsible use ftw!

3/3 My thinking is: - AI can turn a busy person's 50-page doc on an idea into a great pitch - AI cannot just ingest our RFP and produce a great pitch as a response (yet) So the people who had an idea spelled out could use AI to refactor it and do a good job, and the people who didn't couldn't. Moving forward we'll use AI writing detectors but not punish AI writing in isolation -- only when also coupled with a lack of substance, which low-context AI writing is prone to producing.

PSA that we retroactively ran an AI writing detector over all the pitches we got, and found that we advanced human-written pitches ~2.4x as often as AI-generated pitches. 1/3

dandelion chocolate total victory

@DanTVusa @NYC @photopills @SonyAlpha This is insane 10/10 work

Clear skies for a Crescent Moon 4am moonrise. Shot from over 14 miles away from NYC at 600mm #timelapse

@ericadamsfornyc @pangramlabs is this ai?

Honored to accept honorary Albanian citizenship, not just as a personal distinction, but as a reflection of the deep and lasting bond between New York City and the Albanian people. NYC is home to one of the strongest Albanian communities in America, hardworking, proud, and deeply rooted in our city. I accept this honor on their behalf and in recognition of the shared values that connect us: family, resilience, entrepreneurship, and a commitment to community. Thank you to @ediramaal, @BajramBegajAL, and the people of Albania for this meaningful recognition. New York is the global capital, and this only strengthens the bridge between our communities.

(ps we’re hiring dm me!)

working @pangramlabs is pretty great :)

@_ueaj why would the FPR matter at scale when u can just have another agent confirm it

the mythos cope on here is genuinely insane - Yes, I think companies should actually gatekeep huge leaps in cybersec risk. Making something like that GA without extensive testing of filters is bad actually. - The general deference to individual empowerment over the keeping of the commons is a bad trend. We are no longer in the "move fast and break things" phase of AI. We are at scale. Rigor and caution are often lessons learned in blood and it'd be nice to avoid that. - They intend to make it GA, but once they're done testing filters with a new opus. This is good - The open source model thing is cope, they pointed the model to the place in the code where it is, the fpr matters a lot at scale and it would be very high. - If you could actually just make a harness and get the same capabilities, where are all the zero days? I'm sure you could brute force your way to them with more test time compute scaling, but the whole sell of higher intelligence is higher heuristic search, which exponentially reduces costs.