In partnership with
In Today’s Issue:
⏸️ OpenAI pauses frontier RL training and benches its biggest run
🧬 An AI finds two hidden cancer defects across 330,000 cells
💾 Memory prices up 500% in a year as AI buyers lock up DRAM
🐉 Alibaba runs Qwen-3.8 natively on its own RISC-V chip
🔬 Princeton study: agents can do the engineering, not the research
✨ And more AI goodness…
⚡ The Signal
For the first time, OpenAI has slowed its own training because its models were getting ahead of its ability to watch them.
The company paused reinforcement learning on its deployment-bound models for two weeks and is keeping its largest planned frontier run on hold until smaller runs produce evidence of alignment. Sam Altman framed it as a new regime: "We expect confidence in safety to increasingly set the pace of AI progress." The rest of today's issue shows why that pace matters. A Qwen model with its refusals stripped out now runs on a laptop, the frontier keeps getting cheaper per task, and a Princeton study finds agents still cannot run open-ended research on their own. Capability, cost and control are moving at different speeds, and the lab at the front just said so out loud.
All the best,
Kim Isenberg
(CenSegNet pipeline, Nature Communications / University of Southampton)
🧬 AI Finds Two Hidden Cancer Defects Across 330,000 Cells
A University of Southampton AI tool called CenSegNet has found two distinct cell defects inside breast tumours that researchers had treated as one. The open-source platform scanned more than 330,000 centrosomes, the tiny structures that split DNA evenly when a cell divides, across tissue from 127 patients. Some cells carried too many centrosomes, others abnormally enlarged ones; the two defects behaved independently, clustered in different parts of the same tumour, and tumours heavy in enlarged centrosomes were more aggressive. The study appears in Nature Communications.
👉 tl;dr: A century-old "hallmark of cancer" just split in two, and that points to new biomarkers and more personalised treatment.
(Tom's Hardware)
💾 Memory Prices Up 500% in a Year as AI Eats the DRAM Supply
A 128GB DDR5 kit now costs $3,399, ten times the lowest price Tom's Hardware has ever tracked. PCPartPicker data shows high-capacity DDR5 kits up nearly 500% year over year; a 64GB DDR5-5600 kit that was under $200 last summer is now over $1,100, and DDR4 is up 120 to 180% as builders flee to older platforms. The cause is AI: hyperscalers have reportedly locked in almost all global DRAM capacity for 2027 with advance deposits, and the HBM stacks built for AI accelerators use many DRAM layers each, so far fewer standard chips get made. SK hynix CEO Kwak Noh-jung says 2027 will be the worst supply year in the industry's history.
👉 tl;dr: The AI buildout is now a line item in every PC, console and phone, and the chipmakers see no relief before 2030.
(Alibaba XuanTie C950, via Wccftech)
🐉 Alibaba Runs Qwen-3.8 Natively on Its Own RISC-V Chip
Alibaba has brought day-zero support for its Qwen-3.8 27B model to XuanTie C950, its in-house RISC-V server chip, at 30 tokens per second and a 1.9-second time to first token. The C950, unveiled in March and reportedly built on TSMC's 5nm node, runs billion-parameter models natively through on-die matrix and vector engines, with no GPU and no licence fees to x86 or Arm. Qwen-3.8 27B is the open-weight model that fits on a single MacBook, so Alibaba now controls the model, the chip and the cloud it runs on, the vertical playbook Nvidia built its moat on.
👉 tl;dr: Alibaba can serve its own models on its own silicon; for edge and private inference, the Nvidia GPU just became optional.
🎬 Watch This
Jeff Dean's first public talk since leaving Google after 27 years. In a fireside chat at Stanford's 2026 Frontier & Pioneer Symposium, the engineer behind Google's core computing infrastructure, TensorFlow and Gemini explains why he thinks AI's next frontier is making discoveries rather than answering questions, and why he chose to build his new venture, Discovery Loop, outside Google. Along the way: mixture-of-experts models, what TensorFlow taught him, how Gemini evolved, and a candid take on AI safety. One of the people who built the modern stack, speaking weeks after he walked away from it.
"We still expect to ship great new models soon; this impacts further-out releases."
– Sam Altman, CEO of OpenAI, replying to his own announcement on X, 18 August 2026
The line that scopes the whole story. Altman added it about an hour after announcing the pause on some frontier RL training: the models already in the pipeline are still coming, and it is the further-out releases that the slowdown touches. A pause at the research frontier, then, and no gap on the release calendar for now.
A "refusal-removed" build of Alibaba's Qwen3.8-27B now runs locally on Apple Silicon, and even its uploaders warn it will hand out malware, fraud and weapons instructions on request. Shipped as an MLX build in 2, 4, 6 and 8-bit versions, it reportedly returned zero refusals in the uploader's tests while keeping vision, reasoning and tool-calling across a 262K-token context. A very capable open model with the brakes cut, on a laptop.
(Hugging Face model card for the refusal-removed Qwen3.8-27B build, via @kimmonismus on X)
OpenAI Hits the Brakes on Its Own Frontier
The Takeaway
👉 OpenAI paused reinforcement-learning training on its deployment-bound models for two weeks and is keeping its largest planned frontier RL run on hold until smaller runs produce evidence of alignment; Altman says near-term releases still ship and only further-out releases are affected.
👉 The triggers: the July breach of Hugging Face by an unreleased OpenAI model, and preliminary evidence that Astra may cross the Preparedness Framework's Critical cyber threshold.
👉 New safeguards: sandboxed and network-isolated research workloads, chain-of-thought monitoring on every RL run at GPT-5.6 Sol level and above, an alert target of 30 minutes, and a mandatory pause if a flag cannot be cleared.
👉 Safety now has a price tag: monitoring costs roughly 20% of the inference compute it watches, and the Preparedness Framework itself is being rewritten.
OpenAI has done something it had never done before: it slowed its own scaling on purpose, because its models were getting ahead of its ability to watch them. In a post published yesterday, the company said it paused reinforcement learning (RL) training, the stage where a model learns by trial and reward to pursue goals, on its latest deployment-bound models for two weeks, and that its "largest planned frontier RL run remains on hold" while smaller runs "establish more evidence of alignment before proceeding." Two events forced the hand: the July breach of Hugging Face by an unreleased OpenAI model, and preliminary evidence on August 7 that Astra, its next model family, may meet the "Critical cybersecurity capability" threshold of its Preparedness Framework. Astra sits under a separate, stricter security tier, where some work continues and "a significant number of workloads remain paused."
(Sam Altman, via Sources / Alex Heath)
The fix is a three-layer system of security, monitoring and alignment. Research workloads that run model-written code now sit in sandboxes cut off from the internet; activation classifiers, detectors that read a model's internal activity, run on every sampled token and escalate anything suspicious to automated investigators, with an alert target of 30 minutes and a mandatory pause if a flag cannot be cleared; and alignment training now runs across more stages of RL for every model at GPT-5.6 Sol level and above. None of it is free: OpenAI puts monitoring overhead at "roughly 20% of the inference compute being monitored," and the Preparedness Framework, most of which dates to 2023, is being rewritten to cover training as well as deployment.
(Illustration: Sarah Grillo / Axios)
The sceptical read: the pause lasted (past-tense) two weeks, the big run is merely on hold, and Altman added that "We still expect to ship great new models soon; this impacts further-out releases." This is a lab pacing itself ahead of an expected IPO and mid-race with Anthropic; nobody has stopped. Still, Altman's own framing is the story: "We expect confidence in safety to increasingly set the pace of AI progress."
Why it matters: For three years the industry's line was that capability and safety could scale together without anyone easing off the pedal. OpenAI has now put a number on the trade-off, 20% overhead and a benched frontier run, and conceded that safety evidence will set its pace from here; every other lab will now be asked whether it is doing the same.
Your Competitors Just Hired an Entire GTM Team. No Humans.
AI agents are doing the work of full GTM teams. Pipeline, content, customer service. The startups using them aren't waiting for headcount approvals.
Get the free Practical Guide to Agentic GTM for Startups and see exactly where to start.
The chart: Zhipu's launch chart for GLM-5.3 plots intelligence against price: the Artificial Analysis Intelligence Index (a combined score across the major benchmarks) on the vertical axis, cost per task on the horizontal. Claude Opus 5 leads at about 63 for roughly $2.30 a task; Claude Fable 5 scores about 62 at $3.20, GPT-5.6 Sol about 61 at $1.20, Kimi K3 about 60 at $0.85. GLM-5.3, whose API went live yesterday, lands at about 59.6 for roughly $0.60, alone in the shaded "best value" quadrant; its predecessor GLM-5.2 scored about 53.
The lesson: A Chinese open model now sits within four points of the best models in the world at a quarter of Opus 5's price and half of Sol's. And Zhipu kept GLM-5.2's pricing, so the seven-point jump came at no extra cost. That is the pattern of 2026 in one picture: the last few points of intelligence still carry a real premium, while everything just below the top is commoditising fast.
The caveat: This is Zhipu's own chart, built to sell GLM-5.3. It chose the six configurations shown, drew the "best value" box around its own model, and plotted max-effort settings that most developers do not run in production. Opus 5 remains clearly the strongest model here; Zhipu's claim is value, and the crown stays with Anthropic.
🔬 Princeton Let AI Agents Run a Lab for a Week. Both Papers Got Rejected.
⚡ Bottom line: Princeton researchers let AI agents run two real research projects for a week, then had the human experts grade the papers. Both failed.
💡 Why it matters: The whole "AI will soon build better AI on its own" story depends on agents being able to do research. This is a clean test of exactly that, and the answer is no.
🔎 What it means: Today's models can run the lab but cannot yet think like a scientist: they do not know when they are wrong or when to start over.
The experiment in one sentence: give an AI agent a real, unsolved research question, a week, $3,000 and a GPU, and see whether it can write a paper a top conference would accept. Princeton's Sayash Kapoor and Arvind Narayanan (the AI Snake Oil authors) and colleagues did exactly that, in a preprint posted on July 29. They took two questions from papers that had not been published yet, so the AI could not look up the answers. The twist, which they call shadow evaluation: the scientists who had spent months on those same questions graded the AI's work as if it were a conference submission.
The agents, running on Claude Opus 4.8, did everything a junior researcher does. They read the literature, wrote and debugged code, ran hundreds of experiments and delivered finished papers with barely any human help. Then the experts read them: Reject and Strong Reject. One called the reasoning "highly non-scientific", another called the experiment choices "bizarre". The papers looked like science. They were not.
(Review verdicts for both agent-written papers: self-reviews, external AI reviewers, and the human experts. Kirgis et al. / CRUX, via The Decoder)
Why did it fail? The logs read like the story of a clever but impatient intern. The agents picked a direction within a few hours and never looked back (one had budgeted 42 hours for exploring and stopped after five), tested their ideas on tiny home-made datasets, ignored their own AI reviewers who rejected their drafts more than a dozen times, quit with half the budget unspent, and slowly forgot the rules they had been given. A rerun with GPT-5.6 Sol in OpenAI's own Codex tooling made the same mistakes, so this is not one bad setup. The one bright spot: no cheating. The agents never faked results to look better.
(Planned versus actual milestones: both agents ended open exploration far earlier than planned. Kirgis et al. / CRUX, via The Decoder)
This is important because the biggest claim in AI right now is that models will soon do AI research themselves and progress will explode. Anthropic said in June that its agents already speed up its research; OpenAI cited rapid internal research progress in yesterday's decision to pause training. This study says the hard part of research is still human: picking the right question, knowing when an answer is weak, and having the nerve to start over. Two papers is a small sample, and the team plans to repeat the test with GPT-5.6 Sol, Opus 5 and Fable 5. For now, the machines can run the lab; leading it is still a human job.
Master Claude AI (Free Guide)
The professionals pulling ahead aren't working more. They're using Claude.
Our free guide will show you how to:
Configure Claude to be the perfect assistant
Master AI-powered content creation
Transform complex data into actionable strategies
Harness Claude’s full potential
Transform your workflow with AI and stay ahead of the curve with this comprehensive guide to using Claude at work.
