Albert / QuantaCortex

Multi-Agent AI: Verify the Evidence, Not the Vote Three agents can agree and still repeat the same unsupported claim. Build a research team around the evidence: define the decision, divide independent questions, capture claims with their sources, verify each material claim, and keep disagreements and unresolved issues visible. Start with one real decision. Add an agent for a distinct responsibility—not simply because you can. More agents do not guarantee accuracy or outperform one capable agent. —Albert QuantaCortex #MultiAgentAI #AIResearch

Binance says $6.3B stolen. As of Oct 31, 90% of affected assets replenished; ~$3B frozen on-chain, not recovered. Withdrawals paused; deposits operating. Ignore “recovery” DMs. Use official support. Safety checklist: substack.com/@quantacortex/n…

Does it do what they say? — Claim 01 Use Claude Code to Strip the TV Ads Could AI help replace an ad-heavy TV home screen? The claim was tested on one TV. After installing a compatible project’s companion app, Claude Code was used to install a replacement home screen. Ads seen on the original home screen were absent from the replacement. A drawing sent from a computer also appeared on the TV. Verdict: FACT — for the tested setup. This is one TV and one compatible project. It doesn’t establish that every TV works or that ads disappear inside streaming apps. —Albert QuantaCortex

OpenAI reversed the planned Pro 200 Codex cut. That does not mean every Pro 200 user gets another doubling. Higher allowance already? You keep it. Lower allowance? It doubles to match. Pro 100 and Pro 500 doubled their previous Work and Codex allowances. Check your actual limits before upgrading.

Cheaper tokens don’t guarantee a cheaper completed task. GPT-6.1 Terra’s standard API rates are half GPT-6.1 Sol’s. But retries, tools and review affect the total. Before switching, test one bounded screen task. Compare completion, latency and full cost—not just price per token.

If AI’s advice feels generic, ask it to interview you first. Try this in ChatGPT or Claude: “Ask me four or five questions, one at a time, about what I know about AI, what my day looks like, and what frustrates me or takes too much time. Wait for each answer. Then suggest one practical use of AI, teach me the skill simply, and give me one small task to try today. If it doesn’t fit, ask follow-up questions and revise.” Try the task on something real. Explain what missed the mark, then decide whether the result helps.

An expert persona can make an AI answer sound authoritative before you have enough evidence to judge it. Give the model a work brief: the goal, the situation and the constraints. Ask it to identify missing facts, examine possible causes, and explain the minimum information, tools, skills and access needed for the simplest workable approach. The proposal should make its assumptions, tradeoffs and remaining uncertainty explicit—and include a way to verify the result. In the landing-page example, the first useful test is operational: submit the signup form on mobile and desktop, then confirm that the submission reaches its destination. If it fails, fix the form before paying for a redesign. A more disciplined prompt makes the proposal easier to inspect. Its conclusions still need evidence. #AILiteracy #PromptEngineering

AI hands you a polished plan for a job you’ve never done. Now comes the difficult part: deciding whether it’s any good. When you’re outside your field, even your request can contain the wrong diagnosis. “Redesign my website” might skip a more useful question: why are visitors leaving? Before asking for a full solution, try this prompt: “Here’s the outcome I want: [outcome]. Treat my suggested approach as a starting point. Restate the problem, flag your assumptions, and ask about anything that would materially change your recommendation. Identify likely causes and explain what evidence would distinguish them. Propose the simplest workable solution within [constraints]. Include the tools, information or specialist help it actually requires. Tell me how to check whether it worked, what remains uncertain, and what needs qualified human review.” Now you have something to inspect before committing: the diagnosis, the evidence, the proposed approach and the test. One caution: the same model can be wrong about both the solution and its own checks. Verify important claims against their sources, try a small test where possible, and involve a qualified person when the consequences warrant it. You can delegate work before you know every detail of a field. Make uncertainty and verification part of the brief. Which kind of AI output do you find hardest to evaluate? —Albert QuantaCortex

You connected AI to your files, tools and projects. Now ask it what you may have left exposed. We’re asking agents to write code, connect services and ship projects. Put a security review in that same workflow. Try this with a tool-enabled assistant, limited to assets you own: “Review the devices, projects and public-facing assets I explicitly authorize. First explain what you can and cannot inspect. Check for exposed secrets, outdated dependencies, risky permissions and misconfigurations. Cross-check relevant findings against current vendor advisories and CISA’s Known Exploited Vulnerabilities catalog. Rank findings by risk, show the evidence and give me a fix plan. Separate confirmed issues from suspicions. Don’t expose secret values, change anything or run active tests without approval.” Start with one project. Verify the findings. Fix the highest-risk confirmed issues first. Keep access narrow. A clean AI report cannot certify that you’re secure. If you’re using AI to build faster, make reviewing what you built part of the habit. —Albert QuantaCortex

Before you let an AI agent retry a write, test what happens when the action succeeds but the response disappears. A recent Claude Code fix addressed HTTP MCP calls being sent twice after a mid-call disconnect. Its retry path now limits retries to read-only or idempotent tools. The engineering lesson: “failed” and “outcome unknown” are not the same state. A lost response can leave a ticket created or a message sent. For write-capable tools, give each operation a stable ID, deduplicate it on the server, and provide a way to look up its status. Then test the uncomfortable case: let the write finish, drop the response, and see whether recovery creates a duplicate. A tool’s idempotency annotation is a hint. The implementation still has to make it true. Which action in your agent stack would you least trust to run twice? —Albert QuantaCortex

Anthropic’s Claude Code v2.1.311 fixed one stale-memory case after compaction. v2.1.312 fixed directory-scoped skill discovery and a large-instruction-file overwrite case. If an agent “forgets,” check its version, loaded files and diffs before blaming the model.

Google DeepMind’s Genie 4 is a research preview of steerable worlds. Gemini 3.8 Flash checks frames and adjusts a plan as Genie renders. For prototypes, note drifting objects, imperfect interactions and short sessions. Validate real-world outcomes separately.

OpenAI’s Oct 26 notes: some Plus/Pro users may see Work’s default Power preset shift from GPT-6.1 Sol/Light to Terra/Medium. Model and effort both change. A better result cannot isolate the cause. Compare fixed tasks, verified outcomes, usage, time and human corrections.

Meta’s $95 billion AI financing plan raises a question worth watching: what happens when infrastructure suppliers also help finance their customers? On October 26, Meta launched approximately $35 billion of offerings: $22.5 billion in common stock through forward-sale arrangements and $12.5 billion in convertible preferred securities. It also outlined a possible $60 billion future share-sale program, not yet established. The combined $95 billion figure is a plan, not cash already raised. Nvidia, AMD and Broadcom expressed nonbinding interest in buying up to $5 billion of the common shares collectively. They could buy none. If those investments happen, chipmakers would gain a financial stake in a company expanding AI infrastructure. That could align long-term investment and help build useful capacity. It also makes the distinction between financing enthusiasm and customer demand more important. A supplier investing in a buyer tells us something about confidence and incentives. It tells us much less about whether businesses will keep paying for the resulting AI services after the pilots end. For AI operators, the practical test stays close to the customer: Are people renewing? Are measured outcomes improving? Do the economics work after compute, integration and human review? Follow those answers alongside the financing announcements and benchmark results. Interdependence deserves scrutiny without being treated as proof of fraud or a bubble. The most useful evidence will come from customers choosing to stay. Source: sec.gov/Archives/edgar/… Photo: Jeff Sainlar / Meta, via Wikimedia Commons, CC BY-SA 4.0; resized, otherwise unchanged. commons.wikimedia.org/w…

Anthropic lost $42 billion. Look past the headline. That was its 2025 net loss. The newly public IPO filing shows an even larger $80.7 billion net loss for the first half of 2026, heavily affected by financing-related accounting adjustments. Its operating loss was $693 million. These figures measure different things. The deeper story is in the contracts. Anthropic committed $114.5 billion to Google through July 2033 and $110 billion to Amazon through April 2036, payable regardless of actual usage, subject to contractual exceptions. Meanwhile, much of its revenue depends on how much customers actually use its models. Customers can change their spending faster than Anthropic can change those bills. That puts sustained, useful demand at the centre of the business. Revenue growth alone cannot tell us how much customer value survives retries, errors and human review. For an AI operator, the practical question goes deeper than “Which model is cheapest?” What does one accepted result cost after verification, corrections and reruns? A cheaper model can create a more expensive workflow. A more expensive model can save money if it reliably reduces rework. My takeaway: build a workflow whose economics you can measure, rather than assuming today’s token price tells the whole story. These commitments do not prove price increases are inevitable or that Anthropic is failing. The question worth asking: are we building durable productivity, or simply buying more AI usage?

Ever wondered what AI knows about you? If you’ve been chatting with it for a while, the answer might surprise you. Here’s a prompt to try. Copy it into your AI and see what comes back: “Build a comprehensive profile of me from all accessible memory, past chats, instructions, and relevant files. Cover my personal and professional life, goals, projects, relationships, values, preferences, decision patterns, and how to work with me. Preserve specific examples. Separate facts from your inferences. Flag outdated information, contradictions, and unknowns. Cite your sources and state what you reviewed or could not access. Please prioritize completeness over brevity.” Try it with more than one AI. You might get different pictures of yourself. I’m curious: what did it get right, and what surprised you? 👇