Sponsored by
Your AI assistant can save a mistake and carry it into the next task. Anthropic is testing “dreaming” to help Claude check what it remembers, but deciding which lessons to keep brings its own problems.
What's on FIRE 🔥
IN PARTNERSHIP WITH HUBSPOT
Turn AI into Your Income Engine
Ready to transform artificial intelligence from a buzzword into your personal revenue generator
HubSpot’s groundbreaking guide "200+ AI-Powered Income Ideas" is your gateway to financial innovation in the digital age.
Inside you'll discover:
A curated collection of 200+ profitable opportunities spanning content creation, e-commerce, gaming, and emerging digital markets—each vetted for real-world potential
Step-by-step implementation guides designed for beginners, making AI accessible regardless of your technical background
Cutting-edge strategies aligned with current market trends, ensuring your ventures stay ahead of the curve
Download your guide today and unlock a future where artificial intelligence powers your success. Your next income stream is waiting.
AI INSIGHTS
Anthropic just shared how it’s building better memory for Claude, including a process called “dreaming.” The idea: use time between tasks to review past work, fix bad notes, and organize what’s worth keeping. Here’s what stood out:
Claude can check its own memories: During a Pokémon Red test, Claude saved the wrong location for a landmark and kept getting stuck. A dreaming pass checked the original session log, corrected the note, and helped it reach the next city.
Saving useful lessons improved performance: On a database task in CL-Bench, Claude Opus 5.5 scored 0.838 with memory, up from 0.655 without it. The stronger models saved tested rules and warnings they could use on later questions.
A wiki helped Claude connect scattered files: Using roughly 350,000 internal files from Stainless, Anthropic had Claude build linked reference pages. Combining the wiki with search raised its score from 53% to 61%, while saving about 20,000 tokens per question.
Strict memory rules sometimes made things worse: Telling stronger models to “record only failures” cut much of the benefit. Useful memory also needs successful methods and facts the agent has checked.
There’s a catch: dreaming uses extra computing power, and stored notes can still be wrong or go out of date. These are research results, so they don’t mean every Claude chat now runs this process.
For anyone building AI agents, the practical lesson is clear: give the agent a way to check and update what it remembers. A bad note can send the next session down the same wrong path.
PRESENTED BY MARKETBEAT
10 AI Stocks to Lead the Next Decade
AI isn’t a tech trend – it’s a full-blown, multi-trillion dollar race, and 10 companies are already pulling ahead.
These are the innovators driving real revenue, attracting institutional attention, and positioning for massive growth.
Get all 10 tickers in The 10 Best AI Stocks to Own in 2026, free today.
AI SOURCES FROM AI FIRE
1. Free Guide: Gemini Notebook's HUGE 2026 Upgrade: Talk, Research & Learn from Anywhere! Google's most underrated AI tools is now WAY more powerful. It lets you talk to your research, and turn complex information into easy-to-follow learning materials, all from your phone.
2. I've Built 100+ AI Second Brains. Here're the Only 18 Ultimate Tips You Actually Need! Build your AI second brain in just 30 minutes, keep it updated automatically, and make Claude, ChatGPT, and your entire team work smarter with shared context.
3. ChatGPT 2.0: Another HUGE Upgrade: Your AI Assistant Is Now One iMessage Away! And that's just one of the recent updates. Here's what's changing, who can access the new features, and how they could make your daily work easier.
TODAY IN AI
AI HIGHLIGHTS
🧠 Meta is bringing Active Memory & Active Focus to Muse and Meta AI, helping them remember key details, handle problems, and follow changes mid-task.
🔬 Anthropic’s Claude helped map a mysterious protein shell, cutting lab prep from 3 days to 90 minutes. A first 3D model followed 34 hours later.
💰 Anthropic reportedly plans to go public before Thanksgiving at around $1.5 trillion, aiming to raise $80B–$100B despite the tech sell-off, per the Financial Times.
👀 Mira Murati’s Thinking Machines released Inkling 2 Small, matching Google’s Gemini 3.8 Flash on a major benchmark. Try it on Tinker now; open weights will follow safety testing.
🤝 Perplexity launched “teammates,” AI agents that work 24/7, split tasks, and manage other agents inside Computer. Available to Max subscribers at $200/month.
💰 Big AI Fundraising: OpenAI reportedly raised $11B at a $1.25T valuation, with SoftBank, Sequoia Capital, and Andreessen Horowitz among the round’s backers. It now aims to raise up to $20B total, down from its earlier $30B+ target as AI stocks slide.
HOT PAPERS OF THE WEEK
1/ Stanford researchers explore how AI agents can create 2D and 3D graphics
Researchers from Stanford, UC Berkeley, UT Austin, and Cambridge, including Jiajun Wu and Leonidas Guibas, review 300+ papers on AI-powered computer graphics. They show how AI agents can use code, graphics tools, and feedback to create and refine 2D illustrations, 3D models, and animations. Big shift: AI agents could automate more complex design tasks while giving creators greater control over the final results.
2/ CoworkOS brings AI coworkers closer to real-world automation
The CoworkOS Team introduces CoworkOS, a system that lets AI agents collaborate with humans on documents, presentations, websites, and 3D designs while remembering progress across sessions. It can also manage long-running tasks and coordinate multiple agents. In one experiment, it improved Qwen3.6's average benchmark score from 33.19 to 42.95 over five training rounds. Key idea: AI coworkers could handle longer projects with less manual supervision.
3/ Tsinghua and UC Berkeley make open AI models competitive in robotics
Researchers from Tsinghua University and UC Berkeley, including Pieter Abbeel and Ning Ding, introduce OpenRUA, a system that helps open-weight AI models control robots. Using Qwen3.8 Flash-Next, it scores 31.39 on RoboDojo, beating the reported GPT-6 Astra–RoboProbe score of 28.97. Big impact: More researchers could build capable robots using open AI models without relying on expensive proprietary systems.
NEW EMPOWERED AI TOOLS
🎨 ChatGPT Images 2.6 brings faster image generation, precise editing, native GIF creation, and SVG exports to ChatGPT, Codex, and the API.
🛡️ LiveReview uses AI to score pull request risks, auto-approve low-risk code changes, and flag critical issues for human review.
🧠 Quest is a personal AI advisor that learns your values, goals, and personality to help you reflect and make better life decisions.
🎙️ Empacer is a voice-first AI assistant for small businesses that turns spoken requests into customer replies, social posts, invoices, and more.
AI CHART
Misread a lab photo, and you can send a scientist chasing the wrong result. OpenAI and biotech company Ligo built BioBasicsBench to test that everyday work: reading raw experiment data, spotting errors, and knowing when there isn’t enough evidence.
The benchmark puts AI agents through 100 routine wet-lab tasks. Here’s how seven OpenAI models scored at xhigh reasoning effort:
Model | Average score |
|---|---|
GPT-6 Astra | 66.7% |
GPT-6.1 Sol | 62.0% |
GPT-5.6 Sol | 43.8% |
GPT-6 Luna | 38.3% |
GPT-5.5 | 35.4% |
GPT-5.4 | 33.2% |
GPT-5.2 | 31.2% |
These scores measure credit earned across grading criteria, rather than the share of tasks fully solved.
The standout: Astra gained 22.9 percentage points over GPT-5.6 Sol, while producing roughly half as many output tokens per task: 4,704 versus 9,099.
What improved? In one cloning experiment, GPT-5.6 Sol accepted an empty plate as a valid control. Astra spotted that its label named the wrong antibiotic, making that control unreliable.
But the leader still trips over basic checks. Astra chose the wrong colonies in another cloning task and overcounted colonies by about 12% on one plate.
For scientists, the chart points to a more useful lab assistant. The remaining errors show why checking the evidence still needs human oversight before an AI answer drives the next experiment.
Finally, a Monday that doesn’t feel like one.
Imagine a Saturday where you're not catching up on the CRM you didn't touch all week.
HubSpot's new self-updating CRM doesn't wait for you to log back in. Friday's calls are already logged. The deal that needs a nudge Monday morning is already flagged. The automation you set up once keeps running without you thinking about it. None of it needed you over the weekend.
Customers are already seeing it pay off: 45% more deals closed. Finally, you start Monday caught up, not behind.
We read your emails, comments, and poll replies daily
How would you rate today’s newsletter?
Hit reply and say Hello – we'd love to hear from you!
Like what you're reading? Forward it to friends, and they can sign up here.
Cheers,
The AI Fire Team
