Something's broken in AI, and nobody wants to admit it. A multi-agent system just burned through $47,000 in 11 days on tool-call loops nobody caught. GitHub quietly lobotomized Copilot mid-session. And the research is damning: 90% of companies say AI has changed nothing for their productivity. Meanwhile, DeepSeek V4 just made every expensive model look silly.
What's Breaking
The $47,000 Bug: Tool-Call Loops Are Silently Draining Budgets A multi-agent research system looped for 11 days straight — repeating byte-identical tool calls because nothing in its context flagged previous attempts as duplicates. Total damage: $47,000. Another user lost $237 in a single weekend. A third burned $3,472 across 17 failed iterations. The fix? Seven lines of code — fingerprint tracking plus abort on repeat. If you're running agents in production without idempotency guards, you're one bad loop away from a five-figure invoice.
GitHub's Copilot Bait-and-Switch: Models Stripped Mid-Session GitHub paused all new Copilot Pro sign-ups and silently killed Opus 4.5 and 4.6 for existing users mid-session — zero notification, developers lost active coding context. The new $40/mo Pro+ tier ships Opus 4.7 hardcoded to "Medium Thinking" only, locking out the reasoning modes people actually need. Developers are calling it what it is: they were sold unlimited access, then the service got lobotomized when server bills got too high.
InfoWorld, Medium — The Copilot Bait-and-Switch
90% of Firms Report Zero AI Productivity Impact MIT, NBER, PwC, and IDC all converged on the same finding this week. NBER surveyed 6,000 CEOs across four countries: 90% report no measurable AI effect on productivity. MIT economist Anders Humlum found that half of AI usage time goes to managing the AI itself — quality control, bias checks, prompt tweaking. Workers who seemed "20% more productive" were actually 20% less productive overall. The Solow paradox is back, and it's not subtle.
The Register, MIT CSAIL Podcast
Top AI News
DeepSeek V4: Frontier Quality at 98% Lower Cost DeepSeek released V4-Pro and V4-Flash, and the pricing is brutal for incumbents. V4-Pro hits 80.6% on SWE-bench and 93.5% on LiveCodeBench — competitive with GPT-5.5 — at $1.74/M input tokens versus GPT-5.5 Pro's $30/M. V4-Flash comes in at an absurd $0.14/$0.28 per M tokens. Both are MIT-licensed, 1M context, and partly trained on Huawei Ascend chips. The open-weight quality gap isn't narrowing — it just closed.
GPT-5.5 Launches at 2x Price, Folds Codex OpenAI unified coding into the main model with GPT-5.5, hitting 82.7% on Terminal-Bench 2.0. But the pricing stings: $30/$180 per M tokens for Pro, doubling GPT-5.4's cost. When DeepSeek V4-Pro matches performance at 1/17th the price, enterprises have a real decision to make.
Cohere + Aleph Alpha: First Transatlantic AI Merger Cohere and Aleph Alpha are merging with a $600M Series E led by Schwarz Group, creating a sovereign AI champion for governments and regulated industries. It's a direct response to the growing US/China concentration in AI, and a signal that "sovereign AI" is becoming a real market category, not just rhetoric.
Enterprise GPUs Sitting at 5% Utilization Cast AI analyzed 23,000 Kubernetes clusters and found enterprises provision 20x more GPU capacity than they consume. That's ~5% utilization versus a healthy 50% target. Enterprise AI spending hit $37B in 2025, and most of it is treating GPUs as a "strategic hedge" rather than a governed resource. The silent budget killer nobody talks about.
The numbers don't lie: $37B spent on enterprise AI, 5% GPU utilization, 90% report zero productivity gains. Something's fundamentally misaligned.
Papers That Matter
Lite PPO: Two Simple Techniques Beat Complex RL Methods Forget algorithmic complexity — Lite PPO shows that advantage normalization plus token-level loss consistently outperforms GRPO and DAPO on reasoning benchmarks. No architectural changes, no new loss functions, just two well-understood techniques applied correctly. The simplicity-first approach challenges the field's obsession with over-engineering training pipelines.
Progressive Thought Encoding Solves the Memory Bottleneck This one's practical. It encodes intermediate reasoning steps into fixed-size vectors, dropping peak GPU memory from 89% to 59.8% while gaining +27.83% on AIME. If you're training longer reasoning chains under fixed memory budgets — and who isn't — this is immediately applicable.
What This Means For You
The $47K tool-call loop and the Copilot bait-and-switch aren't isolated incidents — they're symptoms of the same disease. AI vendors sold "unlimited" on economics that don't work, and now the bill's coming due. GitHub can't afford to run premium models on flat pricing. Anthropic can't respond to billing tickets while building agents. The compute-cost wall is real, and it's hitting everyone from individual developers to Fortune 500 infrastructure teams.
The 90% zero-ROI stat should be a wake-up call, not a dismissal. The companies getting value from AI aren't the ones chasing the latest model — they're the ones with measurement discipline, idempotency guards, and actual production monitoring. The 7-line fix for tool-call loops exists. GPU utilization dashboards exist. The question isn't whether AI works; it's whether you're building the infrastructure to make it work.
DeepSeek V4's pricing isn't just a competitive shot — it's a forcing function. When frontier quality costs 98% less from an open-weight model, every enterprise needs to justify why they're paying the premium. Some will have good reasons (compliance, latency, support). Many won't. Audit your AI spend this quarter, because the market just gave you a much cheaper alternative.
Written by The AI Architect team at Atobotz