Research done together

Collaborations

Some of these began with a question we asked under someone's paper or post, one began with a message someone sent us. In each, a person who builds their own models, memories or reasoning tools works on a question with us, in the open. Each story has its own subject and its own twists.

In half a minute

Three stories so far, each in a different corner of AI:

Anagha Langhe checked whether running her model in FP8 instead of bf16 changes its answers. One answer in 120 flipped, and when both builds were wrong they picked the same wrong answer 42 times out of 48.
Mark Goodbody wrote decision cases to see whether his reasoning frameworks change what ChatGPT and Claude decide. In a single run per version they did not: the decisions stayed the same, and the models mostly kept a changed claim apart from a changed action on their own.
Mohan Bheemadolu packs an agent's memory into tiny codes with FlyHash. Together we set up a test of whether a smaller memory makes two checking AIs agree on the same wrong answer more often.

The stories

Anagha put the same 120 questions to a bf16 build of her model and to an FP8 build. Only one answer flipped. When both were wrong, they gave the same wrong answer 42 times out of 48.

Does FP8 change a model's answers? One flip in 120

with Anagha Langhe · self-hosted LLM on one NVIDIA L4

Read the story →

Three advisers recommend the school, all from the same two-year-old map. One adviser has today's inspection saying the school's step-free entrance is closed. Both AIs counted the three as one source, with or without Mark's framework in front of them.

Do AI models already keep a claim apart from an action?

with Mark Goodbody · systems engineer, UK · author of Mechanical Ethics and TRACE

Read the story →

When an agent's memory missed the fact, GPT and Claude still gave the same answer 33 times, and 15 of those were wrong. Mohan's FlyHash makes memory much smaller. Does it also make that shared mistake more common?

When an agent's memory is squeezed into short codes, do two AIs agree on the same wrong answer?

with Mohan Bheemadolu · AI agent developer, Redmond · co-author of FlyHash for LLM agent memory

Read the story →

Our note on choosing AI tools for your own work points to top10ai.com's guide to the best AI app builders and to their Claude vs ChatGPT comparison. Now we are talking about a joint comparison of app builders, run through ChatGPT and Claude. A funny coincidence of names: our 2025 project 10think.ai, "the top 10 from 10 AIs", grew into the Poly rankings.

Which AI app builder to pick? A comparison for two AIs

with Milena Josimov · top10ai.com

Visit top10ai.com →

How a collaboration works

You bring a question, a system or data: a model you run, a retriever, a set of cases, a paper you want to stress-test.
We agree what to run and what would count as an answer, before looking at the results.
Both sides run their part. Everything that can be open is open: code, prompts, raw outputs, numbers.
The story goes on a page like these, in plain words, with your name and links, and grows as the work does.

Propose a collaboration

Tell us in a few lines what you are working on, what you would like to find out, and where your code or paper lives. We read everything and answer personally.

[email protected]