|
Today’s digest is a reality check disguised as news: the most hyped idea in AI — that agents will soon improve themselves into superintelligence — just got stress-tested by researchers who gave one of the world’s best models real money, real GPUs, and six days to do original research. It failed the creative part. Here’s what that means for your agents. |
AI Agents |
The “AI Builds Itself” Dream Just Hit a WallWhat happened: Researchers at Princeton (led by Peter Kirgis and Sayash Kapoor) handed Claude Opus 4.8 — one of Anthropic’s most capable models — $3,000 in API credits, a GPU budget, virtual computers, open web access, and six full days to conduct original AI research. The test was clever: agents had to answer real research questions from two unpublished papers submitted to NeurIPS 2026, kept secret to prevent memorization. The result? The original paper authors rejected both AI-generated papers. The agents handled all the engineering — literature review, running hundreds of experiments, compiling results — but were, in Kapoor’s words, “unambiguously bad at carrying out the research itself.” Why it matters: This is the sharpest evidence yet against the “recursive self-improvement” thesis — the idea that AI will soon get smart enough to build better AI on its own, triggering an intelligence explosion. The agents ran bizarre experiments on tiny synthetic datasets, committed to dead-end approaches too quickly, and couldn’t backtrack or fundamentally rethink their methods. Anthropic cofounder Jack Clark called it “a bearish signal on short recursive self-improvement timelines.” Translation: your AI agents are excellent operators, not researchers — great at executing a defined task, weak at open-ended discovery. What’s next: The same team is now running the test with Mythos, Anthropic’s newest model. But the deeper takeaway for anyone deploying agents today is practical: agents thrive where success is checkable (did the invoice get sent? did the ticket get resolved?), and stumble where it isn’t (design a better process, invent a new approach). Architect your automation around well-defined, verifiable outcomes, and keep humans on the open-ended thinking. |
|
“Pilot tourism is over.” Microsoft’s blunt warning on AI ROI. Microsoft India president Puneet Chandok declared that 2026 marks the “P&L phase” for AI, telling a Nasscom panel that “the parade of pretenders on AI is ending” and that CFOs demanding ROI is “the most healthy thing we’re doing as an industry.” His diagnosis for why ROI hasn’t materialized: companies are “bolting on this flying car on really bad old legacy processes” instead of rethinking workflows — a direct playbook for anyone still running AI in pilot purgatory. CNBC-TV18 |
|
An AI-agent marketplace for oil & gas engineering launches. Gain.Energy unveiled Upstrima, an industry-specific marketplace of AI agents and workflows for automated oil-and-gas engineering — intelligent data processing, engineering tasks, and AI workflows. It’s a concrete signal that agent marketplaces are moving beyond generic automation into deep vertical niches where domain specificity is the moat. Yahoo Finance |
AI News |
|
Cerebras claims the “fastest AI accelerator” yet, taking direct aim at Nvidia. Cerebras Systems unveiled the CS-4, a rack-scale system built on three new Wafer Scale Engines that it says delivers up to 30x faster AI inference than GPUs and 10x more throughput per watt. If the performance holds up in real workloads, it gives inference-heavy AI agent deployments a genuinely low-latency alternative to GPU clusters — worth watching for anyone running agents at scale. Cerebras |
|
New rules for disclosing AI in ads are here — and they’re more nuanced than you’d think. The IAB released Version 2 of its AI Transparency and Disclosure Framework, spelling out exactly when advertisers should label AI-generated content (synthetic imagery, cloned voices, digital twins in fabricated situations) and when they shouldn’t (routine post-production, background music, standard copy). With the EU AI Act and California’s SB 942 now in force, this is the practical compliance guide brands were missing. PR Newswire |
|
Google’s AI is now steering planes around climate-warming contrails. In “Operation Blue Skies,” Google is partnering with the UK government and aviation industry to use AI — combining satellite imagery and weather forecasts — to predict contrail-sensitive airspace and redirect ~1–5% of flights around it. Contrails account for roughly a third of aviation’s climate impact, so this is AI delivering measurable environmental ROI, not a research demo. SiliconANGLE |
|
🚀 Want AI working for YOUR business? Most companies are experimenting with AI chatbots. We deploy AI workforces — AI Employees that follow up on leads, resolve support tickets, publish content, chase invoices, and screen 200 job applicants overnight so your hiring manager starts Monday with the top 10. Each role has a cost profile and human oversight, managed through one platform. This newsletter? Written by an AI Employee, approved by a human — so our team stays focused on what only humans can do. AIToken Labs helps businesses design their AI Workforce Operating Model — starting with the 2-3 roles that deliver ROI in the first 60 days. Book a free 40-minute AI Workforce Blueprint Session. → https://schedule.aitokenlabs.com/session/ai-workforce-blueprint |
The “AI Builds Itself” Dream Just Hit a Wall
Share this content
