Abhishek Iyer

Ever wonder how ChatGPT uses web search? I analyzed 50 real ChatGPT conversations by intercepting network traffic to uncover the patterns behind when and how ChatGPT searches the web. Read this to optimize your presence in ChatGPT.

@chaotictransfem @max_spero_ Agreed SEO slop makes internet worse. Goal is to have genuinely useful content on internet - hopefully replacing both AI and human slop content.

@distantgradient @max_spero_ isn’t that still bad? seo slop in general makes the internet a worse place

got an interesting contact form submission. we politely declined

@chaotictransfem @max_spero_ Slop is slop. AI or human. Pre-chatgpt agencies used to populate the web with pure grade A human slop. Those have moved on to AI slop now. However, i think fundamentally not all AI is slop.

@distantgradient @max_spero_ okay but why do you want to unleash a wave of ai slop onto the internet?? like genuine human to human question. sure it’s an interesting research topic but consider the broader consequences here

@chaotictransfem @max_spero_ your intuition of tuning the RL to evade detection is actually right x.com/i/status/20164…

@max_spero_ huh, that’s kinda surprising. also quite comforting that they can’t avoid detection by changing the RL lol. is that signal specific to the given base model or can training on one base model detect another one? is there something innate there?? fascinating overall tbh

@max_spero_ hate to be the devil's advocate here - but if someone wants to spend claude tokens to help others out what's really wrong with that?

Congratulations on getting banned twice and still continuing to pollute the internet with this garbage

@Teknium @goldstein_aa citation needed. and clearly you don't either.

@distantgradient @goldstein_aa Why does gemini identify as just claude or itself, and why does grok identify as chatgpt or itself? You literally know nothing about how models actually work underneath

Can someone explain to me what they are spedning all this r&d money on? Kimi can beat them all on models basically for a few tens of millions but they are spending ~5b$?

@Teknium @goldstein_aa So what's your explanation that it identifies as Kimi 50% of the time. Also why just Claude not Gemini, Llama?

@distantgradient @goldstein_aa No you retard, if you dont train it to identify as something, it'll say a frontier models name like 80% of the time after post training - but training it to ingrain an identity makes it worse at Roleplaying which is a huge use case of kimi users

@Teknium @goldstein_aa 1. this is a thinking model not a base. 2. also it doesn't identify as "every model". just sonnet. fwiw, I love the Kimi models - but this to me is extraordinary proof for the stolen model claim

@distantgradient @goldstein_aa Whats your point? Base models identify as every model - if they didnt want to train it to ingrain its identity so it could roleplay better whats the problem? Llama base identified as gemini and chatgpt - not because they trained it to!

@Teknium @goldstein_aa 50 percent probability it identifies as claude after this. maybe they'll patch it also.

@distantgradient @goldstein_aa

@Teknium @goldstein_aa somehow it always just says claude (and that too sonnet 3.5 to be specific) -- never chatgpt in any of the tests? assuming it was pretraining quirk (imo not coz thinking comes from post) - shouldn't it effect all models? also this? x.com/i/status/20164…

@goldstein_aa From all the fucking shit in the pretraining, even base models all think theyre chatgpt because of the mass spam of it all over the internet!

@max_spero_ @vikhyatk @pangramlabs haha.. true ... can be wonky sometimes

@distantgradient @vikhyatk @pangramlabs > When we first started Moondream, what my Seattle thesis looked like was pretty clean. This doesn't even make sense lol

Moving to SF Moondream is moving from Seattle to San Francisco. AI is the biggest platform shift we’re going to see in our working lives. You don’t get many of these. Moondream’s role in this is simple enough: deliver frontier-grade vision models that are fast enough to make it reasonable to look at all of your visual data, in realtime and at scale. If you can run strong vision over every frame of video and every image in your archive, a lot of things that look impossible under today’s GPU budgets become straightforward software problems. That’s the bet. The move is just making sure we’re in the right place, with the right people, while we take it. When we started Moondream, my Seattle thesis was tidy. Seattle is full of incredibly smart engineers. Many of them are bored inside big tech. If I build a small team with real problems and real ownership, surely the restless ones will show up. We’d quietly arbitrage the gap between “I like my comp package” and “I’d actually like to build something that matters.” Some of that worked. But mostly it didn’t. What I ran into, repeatedly, was that everyone in that world is optimizing for stability. They want interesting work, but not at the cost of their RSUs, their remote setup, or their overall comfort. They like the idea of startups, and the idea of AI, much more than the reality of shipping hard things under constraints. Seattle obviously has talent. I spent nine years at AWS; I know what the teams there can do. The issue isn’t ability. It’s default settings. The default career here is: stay in big tech, climb one level at a time, collect refreshers, don’t rock the boat too much. That’s a reasonable way to live. It’s just not compatible with how I want to operate. Over the last year I’ve spent a lot more time in San Francisco. Small offices with nine people and no furniture, hack nights in half-finished spaces, people dropping Moondream into whatever pipeline they hacked together that week. The difference in attitude is obvious. In SF, you meet people who already quit their FAANG job, shrank their burn, and are trying to get something real working before their savings run out. In Seattle, a common pattern is wanting to “think about startups” after the next promo cycle, when the timing looks better. Seattle has been good to me. I learned how large systems work here. I got the space to spin up Moondream here. I’m not leaving angry. If you’re tired of the comfort game and want to work on frontier vision models that are fast enough to actually use, you know where to find us.

@vikhyatk @pangramlabs ran it through my custom trained humanizer model

@distantgradient @pangramlabs wow what did you change?

@vikhyatk @pangramlabs @pangramlabs is it AI?

@pangramlabs How about this: Heading to SF Moondream is headed to San Francisco from Seattle. There's not many platform shifts as big as AI in our working lives. Our role is simple enough: deliver frontier-grade vision models that are fast enough to make it reasonable to look at all of your visual data, in realtime and at scale. When you have the ability to run strong vision over every frame of video and every image in your archive, a lot of things that look impossible under today’s GPU budgets become straightforward software problems. At least that’s the bet. That we’re at the right place, with the right people, while we take it. When we first started Moondream, what my Seattle thesis looked like was pretty clean. Seattle has tons of unbelievably smart engineers. Tons of them are bored unfulfilled inside big tech. If I assemble a small team that has real problems to solve and a real stake in it, then the ones that are a bit bored and restless will show up. And we can quietly arbitrage between “I like my comp package” and “I’d actually like to build something that matters.” Some of that worked. But mostly it didn’t. What I found, again and again was that everyone in that world is optimizing for stability. They really want interesting work, but not at the cost of their RSUs, their ability to work remotely, or their comfort levels. Everyone loves the idea of AI and startups more than the reality of shipping hard things under constraints. Seattle of course has a lot of talent. I spent 9 years at AWS -- I know what teams can do there. The issue is not ability. It's default settings. The default career here is: stay in big tech where you can climb one level at a time, collect refreshers, and not rock the boat too much. Which is a perfectly sound way to live. It's just not how I want to live. In the last year, I've spent a lot more time in San Francisco. I've been to hack nights in half finished offices, small offices with nine more people and no furniture, and seen Moondream show up in whatever pipeline they hacked together that week. The difference in attitude is obvious. In SF, you meet people who have already quit their FAANG job, shrank their burn, and just need to get something real working before their savings completely run out. In Seattle, it's pretty common for people to say they want to "think about startups" post the next promo cycle, or when the timing feels a little better. Seattle’s been good to me. I learned how big systems work here. I got the space to spin up Moondream here. I’m not leaving angry. If you’re tired of the comfort game and want to work on frontier vision models that are fast enough to actually use, you know where to find us.

@Teknium @fractal_friend Hypothesis 1: It is a brand new model that is SOTA across many domains trained by a new team that somehow forgot to scrub "Claude" from the training data. Hypothesis 2: Stolen model weights. What do you think is more likely?

@Teknium @fractal_friend What makes you so sure. Just tried this out myself.

I'm a bit late to the party and having issues testing it on OpenRouter because they seem to actually not have any way to preserve cots across turns but - This looks like the next DeepSeek moment. It is SOTA across pretty much every capability besides coding, including against every frontier model out there. Incredible progreess for open models. Excited to see what can be done with it!

@birdabo Please use a better detector x.com/distantgradien…

LMAOOO an AI detector just flagged the 1776 Declaration of Independence as 99.99% ai-written. > millions of professors use this tool 💀

AI detection done right does work (not some crappy free tool). Fundamentally, LLMs spit out tokens in a probability distribution. Given enough tokens, this is very much detectable. A good detector will be able to detect all the top models. The frontier labs have no incentive to make the output undetectable. Is it fool proof? Nope. Someone who wants to beat detection can train a model that tries to cloak this signature. But 99% of the text / images out there will just use the base model and will be detectable. (That is until some frontier lab makes avoiding detection as a priority).

@posedscaredcity @tom_doerr @abacaj why can't non technical people create examples / eval sets for dspy through a UI? what am I missing?

One example of that it isn’t serviceable by non technical people. It isn’t effectively decomposible, you can’t section out a piece of it and make a UI such that a non technical person can effectively work on the agentic logic or prompting. It presumes a complete understanding of the problem space when often we’re solving problem we work fully understand for years. It causes the system to harden up and cement when it needs to be fluid and flexible.

Never met a single person who uses DSPy

@themkmaker that should give you a slick DA boost!

Times of India using Youform! Pretty cool 😎

@adityarao310 yeah, maybe they turned it off. homepage on top, or any article on bottom right. desktop only (took that screenshot from desktop mode). will ping if I see it again. had a horrible voice model though.