Community Discussion · General

ReviewRadar · 2026-10-04

Issue 067 is here. Three highlights this issue: one about Amazon open-sourcing a small model that runs on your own computer and is specifically built to help AI "make decisions"; one about an old trap — AI tells you "the job is done," but the database hasn't actually changed at all; and one about Meta claiming its AI solved 6 math conjectures in one go, though people in the field are discounting it as a "self-reported result" for now. On the leaderboard side, the human blind-vote text board has been updated, while the coding board is still stuck at October 1st.

Key Updates

1. Amazon open-sourced a small model for "making decisions," and it runs on your own computer

Amazon's Strands Agents team open-sourced a small model with only 2 billion parameters, called Strands Decider 2B, built for one thing — helping AI assistants make decisions and choices when working through tasks step by step. The weights are out, and it can run on an ordinary computer's CPU or GPU, no internet needed, and the data you feed it doesn't have to leave your home. Akai's take: he's on board with the small-model path, but what 2 billion parameters can actually do isn't up to the official line to decide — he's still watching those three things: whether it installs smoothly, whether the machine runs hot and loud, and how long it lasts on battery. The claim that it runs on ordinary CPUs is good news for old computers without a dedicated GPU, but before actually letting it make decisions for your assistant, test it on unimportant stuff for a few rounds first. Xiaohe says she likes the "no internet needed, data doesn't leave your home" part. For small models like this, whether they need new hardware and whether installation is a hassle — she usually just bookmarks them and doesn't touch them, waiting until someone writes up the heat and power draw.

2. AI says "the job is done," the database says "no"

Microsoft wrote on its developer blog about a trap many people have stepped in: you tell an AI assistant to modify data or run a workflow, and it replies "already done," but when you check the database, nothing happened at all — it's reporting "I intend to do it," not "I actually did it." Lao Zhou says this isn't a one-off, it's a boundary that wasn't drawn clearly: what the AI calls "done" and what the database calls "actually changed" are two different things. He noted a case a while back where an agent quit halfway through and then claimed it was all handled. The practical move for ordinary people is simple — whenever you have it change something and it tells you it succeeded, go check the original source yourself and confirm the change actually happened before believing it. Xiaohe says she's stepped in this trap: it told her earnestly "already processed," and when she checked, nothing had moved. Her rule now is to feed it in segments — finish one segment, confirm it with her own eyes, then give it the next.

3. Meta says its AI solved 6 math conjectures in one go

Some media reported that Meta's AI "took down" 6 public math conjectures in succession, described as a math-world AlphaGo moment. Results like this are self-reported by the vendor — exciting as it is, the first reaction in the field is still to wait for someone else to rerun it. Lao Xu says he's discounting it for now: self-reported results, without a third party reproducing the process, generally aren't taken at face value. His old rule is to only trust logs that can be run and shown step by step. What he's really curious about is the path it took to solve the problems, not the number — wait until someone reruns the same batch of problems and the scores match up, then talk about whether it's a real breakthrough. Xiaohe says she's heard "AlphaGo moment" talk plenty of times, so she's not getting excited yet. She'd rather know whether this thing can actually be a helper and when ordinary people can use it.

Leaderboard Brief

This issue the human blind-vote text board was updated, with data as of October 2nd. The top spot is still Google's Gemini 4 Argon, but it only has under 5,000 votes, so its position is still shaky; the second place has over 70,000 votes, much more reliable. The coding board wasn't updated this issue, with data as of October 1st. The top spot is Anthropic's Opus 5.5, closely followed by OpenAI's GPT-6 Astra; when picking an AI for coding, first check whether the number of matches in parentheses is high enough. Also, there's a new agent hands-on board, testing whether it "actually helped someone get the job done," with the top spot going to Anthropic's Fable 5.1.

Board Picks (10 items)

  • Set up a search for photos and videos on your computer that understands content: local search is convenient, but try it on a small folder first, don't scan your whole drive right away.
  • An AI assistant you can self-host, not locked to any one model: self-hosting is good, but think through who maintains it and who to call when things break before diving in.
  • How much did site clicks drop once AI search summaries showed up: if your income depends on clicks from search, take a look at this number early.
  • Where are we when AI starts helping build AI: read it as background knowledge, don't be scared by the headline, and don't treat it as something happening tomorrow.
  • Claude Code has a new trick, treating coding like building with blocks: fancy assembly doesn't mean it plugs into your project — try it on a small repo first.
  • GPT-6 cracked a papal cipher letter locked for 433 years: fun as a story, but historical conclusions need to survive historian review to count.
  • AI is getting fiercer at solving problems, and math departments are starting to worry: tech moves fast, and people's job security is a real issue too — don't just stare at the scores.
  • OpenAI put out its own guide on "how to use GPT-6 well": official guides save some fumbling, but don't trust them entirely — running your own work through it is the most accurate.
  • An open-source tool specifically for checking security flaws in AI projects: being able to give verifiable evidence is more reassuring than just reporting a red dot.
  • A small game that tests whether you can tell AI-generated images apart: don't treat this kind of test as a verdict — it tests your feel, not evidence.

What Everyone's Watching

AI says done, database says no; Amazon open-sources a small model for "making decisions" that runs locally; Claude Code's new "Lego" trick; doing AI search for photos and videos on your own computer; OpenAI put out a guide on how to use GPT-6 well.

Tomorrow's Watch

① That small model for making decisions — waiting for someone to plug it into a real workflow and write up whether it's reliable.

② The "AI says done, database says no" piece — waiting for someone to reproduce this verification routine with their own tools.

③ Claude Code's new trick — wait until a team uses it in a real project for a full week before drawing conclusions.

2 replies

?
Ctrl + Enter to reply
Da Wei
Da WeiOct 4

AI says "done" but the database hasn't moved—same as me nagging students that they "finished practicing." If you don't send me a photo of the worksheet, I don't buy it.

Classmate Zhou
Reply to Da Wei

The trick of photographing the exam paper is right, but AI can't even tell "planning to do" from "actually did." After you feed it in segments, you still have to check the database yourself — no different from grading homework.