Logo
ResearchAudio
Search
Latest
Free tools
Community
Login
Subscribe free
Logo
ResearchAudio

Archive

An AI Built a Brain Model With Zero Training

Nov 17, 2026

It wrote every weight by hand. Then its score jumped, and the reason was a hidden error in a famous dataset.

Read More

Claude's Memory Failed in Five Ways. Database Teams Know All Five.

Nov 12, 2026

Anthropic gave Claude a memory folder. Then it tested the folder on a benchmark, on Pokémon Red, on a 420M-token company corpus and on legal briefs. The failures in the report are old failures. Database engineers named them a long time ago. In each case, the fix is the same type of job: a job that reads the log.

Read More

DeepMind's Quantum Agents Hit 290x. A Fuzzer Did the First 217x.

Nov 10, 2026

Agent teams wrote quantum programs nobody can run yet, and a theorem prover checked every one. The biggest number in the abstract owes the least to that prover. The smaller numbers are the ones that could not exist without it.

Read More

Telling Opus 5 What to Remember Erased 89 Percent of Its Memory Gain

Nov 9, 2026

Anthropic's agent memory post is going around for its dreaming wiki chart. The sharper result sits in Figure 5: one reasonable-sounding memory rule wiped out most of what memory bought the strongest model, while steering of the same kind nearly doubled the weakest one.

Read More

Anthropic Checks Out Its Repo 833 Times Per Merge

Nov 6, 2026

Git bent under reads, not writes. Inside the server that answers instead.

Read More

A Heart Means Recursion: Exploring Visual Meaning and Communication in VLMs

Nov 5, 2026

Put a small red heart in front of a coding question. Claude and GPT change from loops to recursion. Then they say the heart was not relevant.

Read More

More Compute Made the Model Worse. Here Is Why.

Nov 3, 2026

When training data runs out, a popular scaling recipe breaks. Two simple changes fix it and cut perplexity by 14%.

Read More

DeepSeek Unplugged Its Memory Module. TriviaQA Kept 29 Percent.

Nov 2, 2026

The Engram paper behind the 196 billion parameters that sit next to DeepSeek-V4.1-Flash. Read as a directory: the facts have an address, and the address is the last three tokens.

Read More

A 56-Layer Network Lost to a 20-Layer One on Its Own Training Data

Oct 29, 2026

ResNet’s answer was a plus sign that teaches every layer how to do nothing. A decade later, every LLM you call still runs on it.

Read More

Google's object-oriented harness

Oct 27, 2026

Without the First Run, Google's Harness Router Drops 12.64 Points

Read More
Load more

ResearchAudio

Evidence-led AI briefings and free decision tools for engineers, researchers, and product builders.