We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
Top Stories
Why it matters: AI is now being judged in settings where failure, cost, and clinical outcomes are visible.
AI-assisted personalized cancer therapy posts a positive Phase 3 readout. Merck and Moderna report that intismeran autogene (V940/mRNA-4157) plus Keytruda met the primary recurrence-free-survival and key secondary distant-metastasis-free-survival endpoints in the interim INTerpath-001 analysis of completely resected stage IIB–IV melanoma; overall-survival follow-up continues and the data will be presented at a medical meeting. Moderna says integrated AI algorithms use tumor and blood sequencing to predict up to 34 neoantigens for each patient’s mRNA treatment. The therapy remains investigational, so this is a clinical readout—not an approval.
Codex discloses a destructive-action failure mode and a layered response. Its safety update says GPT-5.6 sometimes misread temporary cleanup, including a malformed command that reused $HOME and could target the actual home directory. The response adds deletion-target checks, fresh temporary directories, high-risk-command escalation, tighter Full access, replay evaluations, and RL tasks/graders; replay tests “substantially reduced” the behavior while preserving normal coding work.
DeepSeek-V4-Pro is now #2 among open models in Agent Arena (#14 overall). Agent Arena reports +6.3% net improvement at a $0.21 median cost per task. It beats V4-Flash (High) on both performance and cost and is strongest on confirmed success (+13.1%) and Bash recovery (+10.9%).
Research & Innovation
Why it matters: Runtime design and feedback quality are becoming as important as the base model.
AgentSysBench finds the agent control plane is a bottleneck. Across ten agentic applications, non-LLM components dominate latency in five; sandbox working sets peak at 28 GB and task latency can diverge 32× across GPU inference, retrieval, and CPU sandboxes. Its design experiments report 29–40% lower latency from task-aware serving, 4.6× lower memory from state offloading, and 35.2% fewer redundant search calls from tool-result caching.
Debate may reduce RLAIF reward hacking. A Google DeepMind AGI Safety thread reports that a weak LLM judge can inflate reward while judge and policy accuracy collapse; debate with an adversarial critic maintained judgments and recovered 45% of the performance gap to RLVR.
Products & Launches
Why it matters: Agents are being packaged as persistent runtimes, not just chat features.
TrueFoundry open-sourced TrueForge under MIT. The vendor-neutral, self-hostable harness handles tool orchestration, context, subagents, approvals, sandboxed execution, and traces. In its 14-task benchmark, it reached roughly the same answers as Claude Managed Agents with about 40% of the tokens and 30% lower cost; routing the work to GLM-5.2 was about 75% cheaper at a similar solve rate.
Cursor’s cloud agents are becoming event-driven operators. They can react to PRs, Slack threads, and schedules, hold a goal over long sessions, and run subagents in isolated VMs to test changes or swarm fixes.
Google is pushing Gemini into student workflows. Eligible college students receive one year of AI Pro in the US or AI Plus in 140+ countries; the new hub adds notebooks, flashcards, quizzes, diagnostic lessons, and progress tracking.
Industry Moves
Why it matters: The commercial layer is consolidating around routing, token economics, and enterprise distribution.
Stripe agreed to acquire OpenRouter. The model gateway routes across 400+ models from more than 80 providers; Stripe says it dynamically selects models by task complexity, price, speed, and reliability, framing token routing as economic infrastructure for AI businesses.
The revenue race is diverging. A WSJ-sourced snapshot relayed in the monitored feed puts OpenAI’s Q2 revenue at $6.7 billion, up 18%, while its operating loss widened to $12.3 billion; Anthropic more than doubled revenue to $11.6 billion, reported a small adjusted operating profit, and surpassed OpenAI for the first time.
Quick Takes
Why it matters: Smaller models, embodied learning, and privacy controls are broadening where AI can run and what it can touch.
- Codex beyond coding: OpenAI says its open-source harness is being embedded in internal apps and operations dashboards; a tax-prep pilot processed 7,000 returns and cut preparation time by about one-third.
- Embodied learning: GeneralistAI claims GEN-1.5 learns a new task in seconds from a demonstration and generalizes, attributing the capability to large-scale physical-data pretraining.
- Open-weight progress: GLM-5.3 scores 71.5% on Terminal-Bench 2.1, ranking second among open-weight models at $0.31 per test; ValsAI says export controls limited evaluation to public benchmarks.
- Private Safety Processing: OpenAI is previewing a system to identify risks across related interactions without giving personnel access to underlying content, while continuing Zero Data Retention for frontier models.