Exponential AI, Still Misread

Dear Readers,

AI is no longer just a story of smarter models—it’s a story of scale, infrastructure, and economics. From Deutsche Bank’s warning that the so-called “AI bubble” is propping up the U.S. economy, to OpenAI’s eye-popping plans for gigawatt-scale compute, the lines between finance, energy, and technology are blurring. What once looked like a software race now increasingly resembles a global infrastructure contest, with the winners determined by who can mobilize the most capital, power, and cooling to sustain the next generation of AI models.

At the same time, venture bets keep flowing into agentic AI startups like AppZen, showing how investors are convinced that AI will remake even the most traditional back-office tasks. The stakes are clear: economies are riding this wave, but they’re also dangerously exposed if it breaks. For now, AI remains both lifeline and risk factor—an industry holding up markets while simultaneously testing their resilience.

In Today’s Issue:

📈 AI progress is still exponential with models tackling more complex tasks.

💥 A Deutsche Bank report suggests the AI "bubble" is currently propping up US economy.

⚡ OpenAI announced plans to deploy 17 gigawatts of compute power

💰 An AI company automating finance workflows secured a $180M investment

And more AI goodness…

✨ And more AI goodness…

All the best,

Researchers resolved a long-standing question in quantum complexity theory (QMA), proving that error reduction in proofs can be pushed to double-exponential precision but no further using black-box methods. Crucially, GPT-5 contributed a key mathematical idea during the proof process, showing that AI can now provide genuine insights in highly advanced mathematics.

Gemini 1.5 Robotics shows how Google is doing everything it can to remain the leader in robotics in the future. Impressive!

It seems as if October will start right off with Gemini 3 and Claude 4.5, so the second week of October will be exciting!

The Takeaway

👉 AI progress tracks exponential growth in task length, not linear gains.

👉 Models are shifting from chat helpers to autonomous coworkers on real deliverables.

👉 Benchmarks on complex, multi-hour tasks show top systems nearing human-level output.

👉 Organizations that redesign workflows for model-native tasks will capture the upside fastest.

Anthropic researcher Julian Schrittweiser has published a fascinating and highly readable article on his blog.

Everyone’s arguing about an “AI plateau,” yet the curves say otherwise. Across agent benchmarks, the time horizon of tasks models can complete - think not trivia, but multi-step work - has kept doubling roughly every seven months. Fresh cross-industry evals now grade models on real deliverables (legal briefs, engineering diagrams, care plans) and early results show top systems nudging toward seasoned human performance in multiple occupations, with some non-OpenAI models leading on select tasks.

Here’s the crux: people judge AI by a chat reply; progress shows up in compound chains of action. That’s where the step-changes land - models stitching longer plans, handling more context, and producing outputs that look like actual work products, not test answers. For builders and operators, this unlocks a near-term shift from “assistant” to “autonomous coworker,” first on hour-scale tasks, then day-long ones.

Why it matters: If these trendlines hold, 2026 is the year knowledge work starts getting re-factored around AI workflows, not just AI helpers. The winners will be teams that redesign processes for model-native deliverables, measurement, and oversight.

Sources:

🔗 https://www.julian.ac/blog/2025/09/27/failing-to-understand-the-exponential-again/

🔗 https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/

🔗 https://openai.com/index/gdpval/

According to a Deutsche Bank note, the ongoing AI investment boom is currently propping up the U.S. economy - without it, the country might already be in recession. However, the warning is loud: this support is fragile, and a crash in AI valuations could drag down broader economic stability. The key takeaway: the AI surge is doing heavy lifting now, but it’s structurally precarious.

Learn More ►