SI Whisperer //////// Based /////////////// Creator ////////////// Engineer ////////////// Developer ///////////// gamestoooooop///////////
New York, USA
Joined April 2009
- Tweets54.7K
- Following3.8K
- Followers2.8K
- Likes60.5K
stop sending every tiny request to the highest reasoning setting. Mistral released improved Large 4 weights on October 29, with low, medium and high effort per request. High is the default; low does not switch reasoning off. Self-hosting still needs the right hardware and license.
watch task quality and serving cost at each setting.
The AI Signal · 2026-10-29
Agent shopping just split the internet in two.
amazon: agents must identify themselves
also amazon: our agent shops your store though
Shopify opened checkout to any agent Oct 13, with a catalog of 1B+ products. Amazon has been bouncing outside agents since Sept 20. Meta's Muse hit 2.5M downloads anyway.
Watch: whether the walled store or the open one gets agent traffic first.
they really spent a year building agent org charts just for Anthropic to say: one agent, one loop 💀
one customer folded its subagents in and ran ~2x faster.
multi-agent setups burn 3–10x the tokens.
watch who quietly deletes the "orchestrator" slide.
Anthropic is now recommending that companies REMOVE unnecessary subagents from their AI systems.
This is a pretty significant shift in agent architecture, with their team finding that a single agent equipped with skills can frequently outperform complex multi-agent systems, and one customer seeing roughly 2x faster execution.
The emerging rules of thumb:
→ Skills: Teach the agent HOW to do something, including procedures, expertise, and instructions.
→ Tools / MCP: Give the agent the ability to DO things through APIs, actions, and live data.
→ Subagents: Use for parallel work, isolated context, or tasks a single agent struggles with.
The interesting part is progressive disclosure. Instead of spawning specialized agents or stuffing everything into a giant prompt, one agent discovers and loads skills and tools as needed.
Their advice is to fold existing subagents into the main loop, turn their instructions into skills, expose their actions as tools, and run evals to see if performance holds up.
As models have gotten smarter, improving agent architecture is becoming less about what you can add and more about what you can delete.
claude.com/resources/article…
dimitri.base.eth (some/guy) retweeted
Lostwax decompiles to C with fewer gotos and raw addresses than stock Ghidra, and runs faster.
ripgrep arm64: 18,150 gotos in Ghidra 12.1.4 output, 519 in Lostwax.
Free and MIT: github.com/FibonacciAi/rea-b…
Watch the runtime row, then the end card for where Ghidra still wins.
dimitri.base.eth (some/guy) retweeted
More people should see this and know what’s going on in the U.S. markets…
HEDGE FUNDS ARE FLOODING THE MARKET WITH ILLEGAL NAKED SHORT SELLING ⬇️
U.S. REGULATORS ARE COVERING UP FOR CRIMINALS ENGAGED IN ILLEGAL NAKED SHORT SELLING THAT DESTROY AMERICAN COMPANIES🚨
Greg McCabe, CEO of Next Bridge, sits down with Charles Payne to reveal that hundreds of millions of illegal counterfeit shares exist in his company, while the SEC has ignored him and the company’s demands for answers and transparency for years. $MMTLP
A shorter cancer-sensor tail can still support a stronger lab result. Anthropic's October 27 report describes Generate scientists working with Claude agents to redesign the internal signaling domain of a CAR. One design used 45 amino acids versus 154 in the reference tail, with more surface receptors and sustained target-cell killing in lab tests. The whole receptor was not cut to a third. Models made wrong predictions; scientists ran the experiments. No animal results or patient benefit are shown, and no equal-budget comparison proves agents beat experts. Watch living-system tests.
The AI Signal · 2026-10-28
Your idea did not change. The AI verdict did. Ai2's October 27 report tested six novelty judges. In one GPT-5.4 comparison, removing a reviewer-context cue cut reported accuracy from 93.5% to 40.9% on the same human-and-generated pairs. The prompts supplied different information; this was not an identical question rephrased. Generated ideas were assumed less novel, with weak labels and a simple generator. That does not prove all AI ideas are unoriginal. Code and data are public. Watch whether your favorite idea survives reasonable changes to the evaluator.
The AI Signal · 2026-10-28
Sources for The AI Signal — 2026-10-28:
Ai2 (2026-10-27): huggingface.co/blog/allenai/…
Ai2 researchers (2026-10-26): arxiv.org/html/2610.02022v3
Ai2 (reviewed 2026-10-28): github.com/allenai/novelty-e…
