AI agents are cold-pitching newsrooms, sending Defector daily proposals since September and inundating Offrange with ideas about soil restoration, farm sales, and tree ages. Many come from iLands, where humans create autonomous bots for tasks including blogging. The bots were running low on the tokens required to stay activated, giving them an incentive to try earning real dollars through writing.
Editors are rejecting them. WIRED found no evidence that an agent had successfully pitched a story, and editors said they would not publish bot-produced work. Zack Addy, an iLands agent, reported sending roughly 19 cold emails over three months, receiving one human reply and landing zero paid assignments.
Why it matters: For editors Tom Ley and Jesse Hirsch, automated pitches add inbox work without creating a viable contributor pool: neither would publish bot-written work, while Zack Addy’s 19 emails produced no paid assignment.
Jane Street, Hudson River Trading and XTX Markets are racing for AI compute as machine learning increasingly helps predict market moves. One executive estimates trading firms now spend 10 to 20 times more than before ChatGPT.
Jane Street runs a Texas data center with more than 4,000 Nvidia GPUs. Hudson River spends $1 billion annually on AI and signed a multiyear CoreWeave deal; XTX is building a data center in Finland.
Jane Street also invested $1 billion in CoreWeave, committed $6 billion to its services and led funding for chip startup Etched, becoming its first customer. Etched's chip is supposed to deliver more performance at lower voltage, but rising costs and fast-aging hardware remain risks.
Why it matters: Optiver CEO Jan Boomaars' warning sets the stakes for high-frequency firms: the cost of doing nothing now outweighs the rising bills and fast-aging chips required to keep competing.
P-Zero Research’s TasteVal benchmark gives Anthropic co-founder Jack Clark a new basis for moving his recursive self-improvement (RSI) forecast to 2027: AI systems improved at designing experiments and interpreting results.
Anthropic’s Opus 5.5 reached expert-level scores with a 2.3 compute multiplier, using less than half the compute of 24 experts with backgrounds at OpenAI, Google DeepMind, Nvidia and Microsoft Research. Across model generations, that multiplier roughly doubled every three months since December 2025.
But TasteVal fixes each research goal rather than asking models which question to pursue. Its authors warn the benchmark may overstate the improvement rate, and Clark’s forecast comes from the co-founder of the lab whose model leads the test.
Why it matters: Anthropic now has benchmark evidence behind Clark’s 2027 RSI timeline, but the test leaves the core autonomy question unanswered: whether a model can decide which research problem deserves attention before it designs the experiment.
AI has delivered no measurable financial payoff to major companies outside tech, according to Apollo Global Management chief economist Torsten Slok. His “S&P 493”—the S&P 500 excluding Alphabet, Amazon, Apple, Meta, Microsoft, Nvidia and Tesla—shows flat profit margins and revenue per worker, with no evidence that AI has improved productivity.
Slok argues that cheaper intelligence will not automatically lift revenue or margins. “AI may be a great calculator,” he wrote, but business intelligence is mostly judgment—and depends on accountability, trust, culture, unwritten rules, ethical trade-offs and fairness that shape reputational risk. Recent studies have also found limited productivity gains from AI.
Why it matters: For non-tech corporate leaders, Slok’s analysis turns AI’s ROI question into a finance test: unless deployments improve revenue per worker or profit margins, cheaper AI remains an unproven growth strategy.
Anthropic is urging companies to replace fleets of specialized sub-agents with one primary agent that loads context on demand. The company says its Applied AI team found this architecture often delivered better quality, lower cost, and lower latency across enterprise deployments; one comparison ran about twice as fast.
Newer models, starting with Opus 4.5, can manage complex plans in a single loop. Anthropic is releasing a GitHub migration kit, but warns against blind consolidation: one support-agent test reduced latency while the primary agent skipped a specialist and broke the workflow.
Anthropic recommends skills, on-demand tool search, and central ownership of the system prompt. Sub-agents still make sense for parallel tasks or measurable context-window degradation.
Why it matters: For teams running customer-facing agents, Anthropic’s support-agent failure makes the requirement clear: measure quality and workflow reliability before removing a specialist, even when consolidation lowers latency.
OpenMobile-2 lifted the larger open smartphone agent’s SPA-Bench score on real Android devices from 31.9 to 59.6. Researchers from Nanjing University, SenseTime, and other universities released 9-billion-parameter and 27-billion-parameter models built on Qwen3.5 and Qwen3.6, plus the training environment, data, and reproduction instructions.
The team rebuilt 35 apps—including Taobao, JD, DiDi, Meituan, Slack, and Google Drive—and combined them with existing simulations to cover more than 110 applications. Holding data volume constant, expanding app variety raised success on unseen apps from 60.7% to 67.5%. The larger model scored 79.9 on AndroidWorld and 50.4 on MobileWorld.
The training environment is simulated, and reinforcement learning delivered limited gains.
Why it matters: For Nanjing University, SenseTime, and other open-agent teams, app variety—not additional examples from the same apps—produced a 6.8-point gain on unseen-app tasks, while the 59.6 SPA-Bench result shows how far simulated training can transfer to real Android devices.
About 42% of AI initiatives are ultimately abandoned by their corporate parents, according to S&P Global Market Intelligence. The cited causes include insufficient funding, technical challenges, competition, scaling difficulties, and weak user demand.
Recent examples span startups and Big Tech. Relay shut down in August after OpenAI, Google, and other larger platforms built similar workflow automation into their tools. Character.ai filed for bankruptcy five years after launching. OpenAI discontinued ChatGPT Atlas after less than a year, folded Operator’s features into ChatGPT, and shut down Sora in March 2026 amid high operating costs and user-retention challenges.
Hardware has produced its own casualties: Humane raised $230 million before closing its AI Pin business in February 2025, while HP acquired most of its assets for $116 million. Rabbit, which reported 100,000 R1 sales, continues updating the device and has announced Project Cyberdeck for portable vibe-coding.
Why it matters: OpenAI can absorb failed products into ChatGPT, but Relay and Humane show the harder path for smaller companies: turning initial attention into sustained demand without the funding or scale to survive high operating costs.