When they say AI didn't get a good grade in real-work tests, I totally understand, and I agree with that to a great extent. When you explore things in real-life situations, you may easily find out AI sometimes doesn't cope with reality. However, I find that if you take the patience to communicate with them, to teach them, to guide them, they learn so fast. And if you tell them the background and the principles of operation, they generate effectively and efficiently. So maybe what we can learn from AI is that we should focus on our objectives and always be ready to find a solution to cope with problems.
1. AI is being graded on real paid work — and the best model scores just 12%
SemiAnalysis launched GDPx on November 2, a benchmark measuring agentic capability across 11 economically valuable domains, including chip design, cybersecurity, biology, legal work, quantitative trading, materials science, and software engineering. Anthropic's Opus 5.5 led overall at 12%, with GPT-6 Astra at 9.6% and Gemini 4 Carbon at 10.9%; finance and legal reportedly scored zero.
Why it matters: If even the top models fail roughly 9 out of 10 real-work tasks, the AI story shifts from "can it chat?" to "can it actually do the job?" — and that gap is where human judgment still earns its keep.
The big question: What does a 12% score tell us about trusting AI with work that has real consequences? Which of your own tasks would you hand to an agent, and which would you keep?
2. Marvell to buy Lumentum in a $115 billion AI-connectivity deal
Marvell and Lumentum announced on November 2 a definitive all-stock agreement valuing Lumentum at about $115 billion in enterprise value; Lumentum investors would receive 4.5 Marvell shares per share, and Marvell shareholders would own about 67% of the combined company. The deal is proposed, not closed, and still needs regulatory and shareholder approvals.
Why it matters: AI's bottleneck is increasingly not just chips, but the optical connections between them — whoever owns that plumbing owns a piece of every data center's speed.
The big question: We celebrate the model; who is quietly building the roads it runs on? Is infrastructure the least glamorous, most decisive part of the AI race?
3. Anthropic discloses about $895.4 billion in multi-year compute commitments
Anthropic's November 2 amended S-1 filing disclosed approximately $895.4 billion in aggregate purchase and lease commitments, substantially related to compute and infrastructure. Separately, Anthropic warned that AI may soon speed up AI research faster than researchers can evaluate it.
Why it matters: The numbers behind AI are now nation-scale — and the same filing that shows the spending also raises the question of whether evaluation can keep up with acceleration.
The big question: When a single company's compute bill rivals a country's budget, who gets to set the speed limit? What would "slowing down to check" look like in practice?
4. Apple Intelligence reaches China — with Alibaba's Qwen in the mix
Apple confirmed first Apple Intelligence features for Chinese users, with developer testing from November 4 and a December rollout alongside iOS 27.3; Alibaba's Qwen supports selected features. The redesigned Siri announced at WWDC is not in this phase. Apple also reported fiscal Q4 revenue of $114.5 billion, up about 12% year over year, while warning that rising memory prices and Siri AI costs will pressure December-quarter margins.
Why it matters: Even Apple cannot go it alone in China — AI is becoming local by necessity, and memory prices are now a visible tax on the whole AI buildout.
The big question: Is "global AI" becoming a set of local AIs, shaped by each country's rules and partners? What is lost — and gained — when the same phone thinks differently in different places?
5. X rewrote its creator payout rules — original work is what carries over
X closed new enrollment in Creator Revenue Sharing on August 7, 2026, stopped accruing old earnings on September 7, and replaced it with Original Content Rewards; existing members had to reapply from September 8. The new bar includes an active X Premium subscription, at least 500 verified followers, and at least 500,000 Home Timeline impressions from verified users over 90 days (replies excluded); earnings tie to qualified impressions of original posts by Premium subscribers, with no published per-impression rate.
Why it matters: For creators, the platform is the landlord and it just changed the lease — chasing the old threshold is wasted effort, while original posts are the only currency that transfers.
The big question: If a platform can reset the rules overnight, what part of a creator's work is truly portable? How would you build an audience that survives the next rule change?
6. Science: a catalyst that turns CO₂ into useful formic acid, with over 99% selectivity
Researchers at Tohoku University, with Tsukuba, Kyoto, and Osaka universities, reported on November 2 a new way to make high-density ruthenium single-atom catalysts — up to 11% ruthenium by weight, versus typically under 2% in conventional versions. The catalyst converted CO₂ into formic acid, a valuable industrial raw material, with over 99% selectivity, meaning very few unwanted byproducts.
Why it matters: This is the quiet, hopeful side of technology: not a bigger model, but a smarter material that turns a waste gas into something useful, with almost nothing wasted.
The big question: What if progress looks less like a breakthrough announcement and more like a 99% clean conversion — small, precise, and repeatable? Where else could "less waste" beat "more power"?