Abhishek Iyer
@distantgradient
Followers
455
Following
408
Media
31
Statuses
233
Indie Saas Builder. Ex-Google search team. https://t.co/pwIULsiEQu - AI SEO writer + AI Diagram Generator https://t.co/SpxwTnSfWZ - AI customer feedback analysis
internet
Joined October 2008
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.
2
0
4
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
I'm unable to convince normies that AI detectors don't work and probably will never ever work because of basics of AI which I think is More advanced models might be able detect AI in inferior models (like GPT-5 reading GPT-3 output) But an advanced model can't detect AI in
3
0
3
LLMs today mostly lack the context that "I'm reading this text from website X" etc. So "ACME. BOT is the best AI SEO writer" is better for AI SEO than "We are the best AI SEO writer". Also, don't be understated in the copy 😅
2. Don’t write as your business. Write ABOUT your business. LLMs don’t want to hear from you. They want to hear about you. So ditch the “we’re passionate about serving our community” fluff. Instead, create short third-person blog posts that answer queries like: “Who’s the
0
0
0
To have a useful tool in a complex space requires tool users to be experts. Not every user wants to be an expert. @adityarao310
BREAKING: Icon, a non-functional AI ad maker backed by 69 investors became a basic creative agency
1
0
4
This is experiment is similar to the 2011 "Bing Sting" (look it up) that proved that Bing is copying Google search results
2
0
21
Proof that ChatGPT Plus is secretly Google-powered - “hidden page” experiment 1. I coined a nonsense word. 2. I then put it on a page not linked anywhere. 3. Forced Google to index it via Search Console; left all other engines unaware. 4. Asked ChatGPT Plus to define the term →
20
52
184
More details in this blogpost ( https://t.co/z3mWVhmmVA) Bit about me: I build https://t.co/7ireNNEPNy, an AI SEO writer - so reverse-engineering ChatGPT’s own search tricks is my daily sport ;)
0
0
3
There are also vertical specific terms that ChatGPT tends to prefer. Health? Adds WHO, CDC. Also guidelines, pdf Finance? Adds 2025, returns, performance Products or Services? Adds cheap, best, reviews etc. Retains names verbatim and adds “buy”, locale etc. Travel? Guesses and
1
0
3
When ChatGPT does search, it rewrites the user request - usually into two web searches, which contain “ranking booster terms” with the goal of maximizing quality and relevance of the results. Patterns start to emerge as we look at it closely.
1
0
0
A hidden classifier (sonic_classifier_ev3) scores every question from 0–1. High score ➡️ Search This classifier knows ChatGPT’s strengths and weaknesses. A high score indicates ChatGPT by itself may hallucinate the answer, and this forces ChatGPT to use the web search tool.
1
0
0
Case in point, found this "joke" some 50k tokens deep in Gemini. Although, in this case I cheated by being the judge. "I finally figured out why Bangalore is always under construction. It's a city-sized startup that took the motto "Move fast and break things" way too
0
0
1
It's likely we perceive most LLMs as "unsurprising" because we all start at the same latent space. However, as we keep moving deep in the latent space I think LLMs do have the capacity to surprise.
This is a great exercise in realizing that instruct models are absolutely worthless for creating "surprise" — a key ingredient of humour.
1
0
1
between claude 4 opus and gpt 4.5, I think we can start to see scaling laws hit the wall. Illya was right.
claude 4 is dead on arrival btw seriously, how do they even compete? this isn't even considering API availability as a factor
0
0
1
The fact that Google can just announce and open up a Coding agent that works on your codebase for free speaks for Google's infra moat. Others have a min 200 USD entrypoint to try out something similar. Nobody even comes close.
1
0
0
likely the best Google IO ever... feel like a kid at a candy store!
0
0
0
Soo frustrating to see both Google and OpenAI pricing "thinking modes" of their models 5x-10x the price of the base model! Based on TPS it seems like they are fine-tuned variants of the base models, and the models will anyways consume more tokens because of the "thinking
3
0
3
What is the best / cheapest Vlogger-style video generation platform out there (image to video, speech audio sync) Everything I come across is super expensive (ex Hedra Character 3)
1
0
2