In partnership with
Welcome back to The AI Field. Meta's Muse can now build you a small app from a description and keep it up to date. Anthropic explained how Claude cleans up its own memory while it is idle, and a research startup says an AI system proved a quantum computing conjecture that had stood since 2008. A weekend issue about AI that remembers, organizes and checks its own work.
Inside: three stories, five tools, three quick hits, the most-viewed AI posts on X, and a five-minute way to have Muse hunt down a Marketplace deal.
In today's AI Field:
Muse builds custom planners and trackers you describe
Claude “dreams” to correct and organize its memory
An AI system proves a quantum conjecture open since 2008
Plus: five AI tools, three quick hits, today's viral AI posts on X, and a five-minute Marketplace deal exercise.
Read time: 5 minutes
Master Claude AI (Free Guide)
The professionals pulling ahead aren't working more. They're using Claude.
Our free guide will show you how to:
Configure Claude to be the perfect assistant
Master AI-powered content creation
Transform complex data into actionable strategies
Harness Claude’s full potential
Transform your workflow with AI and stay ahead of the curve with this comprehensive guide to using Claude at work.
LATEST DEVELOPMENTS
META · MUSE
Watch the Spaces demo (90s video). Source: Meta
The AI Field: Meta's Muse app can now build “spaces”: custom planners, trackers, dashboards or games that you describe in chat. Each space lives inside the app, and you can keep asking Muse to change it as you use it.
The details:
Meta's examples include a holiday gift tracker, a school dashboard that finds your kids' events in school emails and adds them to your calendar, a small-business order tracker, a freelance client hub and a trip planner.
Connect the apps you use and a space updates itself in the background, so new concert dates, price changes or savings progress show up without you asking. You can check things off, add notes and edit right in the space.
You choose where spaces live: the sidebar, the main tabs, hidden, or scheduled to appear at set times. Spaces are available everywhere Muse is, and Meta says it is exploring ways to share them with others.
Meta is also rolling out specialized models trained with Muse Wonder. In Muse, it says they show a 2.4x reduction in forgetting goal-critical information and a 67% improvement on everyday tasks where unexpected issues come up.
Why it matters: The useful change is that a chat result now sticks around. Instead of asking for the same summary every week, you get a page that keeps itself current. The test is the connections: a space is only as good as the inboxes and calendars you let Muse read, so start with one low-stakes tracker and check what it pulls in before you connect anything sensitive.
For the model behind the upgrade, see our issue on Meta's Muse Wonder launch.
TRY THIS TODAY · 5 MINUTES
Find a Marketplace deal with Muse
Maxym had been hunting a PS5 Pro. He says Muse found a Facebook Marketplace listing for $750 with two controllers, a charging station and an extra 2 TB SSD, and he bought it. The useful pattern: let the agent watch new listings and draft the first message, while you decide whether the seller is real. Muse is free to try in the US, Canada, Japan and Singapore.
1. Open Muse and decide on the item, your ZIP code, your max price and the one thing the listing must include.
2. Paste this prompt and fill in the brackets:
"Watch Facebook Marketplace near [ZIP] for a [PS5 PRO / ITEM] under [MAX PRICE], preferably with [ACCESSORIES]. When a new listing matches, show me the price, location, seller join date, photos and the listing link. Draft a short first message asking whether it is still available and whether [MUST-HAVE] is included. Do not pay, share my address or agree to meet until I approve."
3. Marketplace deals move fast and attract scams. Never pay off-platform, meet in a public place or use the site's checkout when you can, and confirm the serial number and condition in person.
TRENDING AI TOOLS
Five tools to try this weekend. Click a name to open it, or browse more in our
👉 AI tools directory.
⚙️ Make: Running Meta ads? Make shows how to calculate each order's net profit and send that to Meta's Conversions API instead of revenue, so your ads chase buyers who actually make you money. Free plan, no card needed.
🧱 Base44: Your app can now open a real browser in the cloud to collect prices or event details, fill in forms, or check that your live app works (Builder plan and up). Free plan to start.
🧠 Genspark: Tired of repeating yourself? SecondBrain pulls your emails, meetings, chats and docs into one memory, so Genspark's Super Agent already knows your work. Free daily credits to try Genspark.
🧑💻 T3 Code: Run Claude Code, Codex, Cursor, Antigravity and other coding agents from one open-source app, using your own subscriptions. Also on iOS and Android.
🌬️ Reynolds: Describe a fluid flow in plain words, and Standard Physics' agent writes the simulation case, runs a real solver and hands back the figures and numbers. Open source, free with a work email.
ANTHROPIC · AGENT MEMORY
Source: Anthropic
The AI Field: In a new engineering post, Anthropic explains how Claude can use idle time to “dream”: reread its past session logs, fix wrong memory notes and build new ones, like a linked wiki, so later sessions find what they need faster.
The details:
In a Pokémon Red test, Claude Sonnet 3.7 wrote 413 memory files across 225 folders and earned one badge. Claude Fable 5 kept nine files in two folders plus one global.md dashboard and earned four.
Over 128 dreaming sessions on about 350,000 internal files from Stainless, a company Anthropic acquired, answers from the wiki alone rose from 13% to 48% of graded criteria. With search plus the wiki, Claude scored 61% versus 53% with search alone, and used about 20,000 fewer tokens per question.
Rigid rules backfired: telling Opus 5 and Opus 4.7 to “record only failures” cost them much of the boost memory gave them. Claude Code's auto memory loads only the first 200 lines or 25 KB of MEMORY.md into each session.
Why it matters: If you keep a memory or instructions file for an AI assistant, keep it short and correct rather than complete. A bloated memory slowed Claude down, and one wrong note sent it back down the same wrong route three times. Every few weeks, review what your assistant has saved and delete what is stale.
Create work your clients can rely on
With Lovable, you'll have a working version of your client's site ready for the first meeting. Plus:
Launch on your client's own domain
Build sites that get found on Google
Catch bugs and errors with project monitoring
Run every client site on one plan
Discount of $5 applies to first month of any paid plan. You need to sign up via this webpage in order to claim the discount. The discount applies to new and returning users without an active paid plan. Terms apply.
PHYSICAL SUPERINTELLIGENCE · MATH
Source: Physical Superintelligence
The AI Field: Physical Superintelligence (PSI) says a research system it built, running a non-public Anthropic model, proved the Aaronson-Ambainis conjecture. It is a central open problem about when quantum computers can be exponentially faster than ordinary ones.
The details:
The system had a complete proof on the third day of its run, and Lean, a proof assistant in which a computer checks every step, accepted it. The system later improved the result twice.
The result: on yes-or-no questions about random inputs, a quantum computer cannot be exponentially faster unless the input has a rare pattern. Shor's famous factoring algorithm works because its problem has that kind of structure.
Stanford researchers who proved a special case in August are reading the full proof and called it “a huge result.” PSI researcher David Turturean led the work and is responsible for the paper's correctness.
Why it matters: A machine-checked proof is a different kind of claim than a chatbot answer, because Lean verifies every step. The human role here was picking the problem, auditing the proof and owning the result. That is a useful template for any team putting agents on hard problems.
QUICK HITS
🔐 Google DeepMind's AI agents find weaknesses in old RSA encryption: Guided by humans, the agents found two weaknesses in legacy 1024-bit RSA, including a way to cut the cost of factoring each extra key in a batch by about 30x. DeepMind says keys of 2048 bits or more remain secure (X post).
🧪 OpenAI launches BioBasicsBench: 100 routine wet-lab tasks built with biotech company Ligo. GPT-6 Astra scored 66.7% versus 43.8% for GPT-5.6 Sol at the highest reasoning setting tested on both.
🛡️ AI labs get a shared way to rate cyber jailbreaks: The Frontier Model Forum's framework scores them on capability gain, scope, ease of deployment and ease of discovery. Public comments are open until December 21.
TRENDING ON X
That's it for today!
Reply and show us how you use AI this weekend. The useful ones may show up in a future issue.
See you soon,
Olle Hellman | Founder of The AI Field