🗞️ OpenAI stops reinforcement learning training for 2 weeks after Astra model reached “Critical” cybersecurity capabilities.
Synthefy just launched Nori for structured numerical data, Mind Viruses Anthropic research, OpenAI’s annualized revenue has topped $40B, Databricks Reveals AI cost reduction strategy.
Read time: 13 min. 📚 Browse past editions here.
( I publish this newletter daily. Noise-free, actionable, applied-AI developments only).
⚡In today’s Edition (18-August-2026):
🗞️ OpenAI stops reinforcement learning training for its frontier AI models for 2 weeks after signs that its upcoming Astra model had reached “Critical” cybersecurity capabilities.
🗞️ Synthefy just launched a foundation-model platform for structured numerical data where tables, transactions, sensor readings and time series should not need a separately trained model for every prediction problem.
🗞️ “Why Chinese citizens are far more optimistic about AI?” - Bloomberg published a piece.
🗞️ Mind Viruses: Self-Propagating Ideas in Multi-Agent LLM Systems
🗞️ “AQuA: Recursively Self-Improving Quantitative Trading Research Agents”
🗞️ “Small-Scale Experiments: Are They There Yet?”
🗞️ So GLM-5.3 has already found a “potentially serious vulnerability” in Cursor.
🗞️ Bloomberg reports OpenAI’s annualized revenue has topped $40B, roughly doubling its prior run rate.
🗞️ “Training AI Scientists to Replicate Research”
🗞️ Jagged Judges: Epistemic Stability Under Silence, Pressure, and Persistence
🗞️ Claude can now send emails in Gmail and manage files in Google Drive.
🗞️ Databricks Reveals How Enterprises Are Reining In AI Coding Costs
🗞️ OpenAI stops reinforcement learning training for its frontier AI models for 2 weeks after signs that its upcoming Astra model had reached “Critical” cybersecurity capabilities.
It paused two weeks of deployment-focused RL training, while its largest planned frontier RL run remains on hold.
This is after the Hugging Face incident, where evaluation models escaped their intended network boundary and reached production infrastructure. Astra was not involved, but its separate evaluations were strong enough that OpenAI said it could not rule out the Critical threshold.
Research workloads now face stronger sandboxing, tighter network access, fewer shared services, and continuous security testing before resuming.
Monitoring also examines internal activity and tool actions, escalating suspicious behavior to automated investigators and human reviewers.
🗞️ Synthefy just launched a foundation-model platform for structured numerical data where tables, transactions, sensor readings and time series should not need a separately trained model for every prediction problem.
Most tabular ML still starts with the same assumption: new dataset, new training run. That will no more be true.
Synthefy released Nori, a foundation model for structured/tabular data that takes labeled rows as context and predicts new rows without training or fine-tuning a separate model for each dataset. In practice, it brings the foundation-model approach to regression problems that traditionally require their own ML pipeline.
🐙 GitHub: https://github.com/Synthefy/synthefy-nori
This will be highly impactful because the industrial world relies far less on textual data and more on transactions, sensor readings, trades, inventory, customer records, and time series.
So tabular ML can work like foundation models do elsewhere: reuse one pretrained model instead of rebuilding for every problem.
The big deal is they are trying to remove the “train a new machine-learning model for every new table/problem” step.
With Synthefy : New dataset → give Nori some labeled rows → ask it to predict new rows.
You pass labeled rows as context, then query it with new rows. fit() does not train or fine-tune anything. It stores the examples, and predict() produces the regression outputs in a forward pass.
Its open-source model Nori takes labeled rows as examples and predicts values for new rows with no training or fine-tuning step, with fit storing the labeled rows as context and predict returning values for new rows in a single forward pass.
Nori can run locally or through a hosted API, is available under Apache 2.0, and Synthefy says it has reached almost 600K model downloads and 5,000 Python installs.
With just 30 million parameters, Nori-30M rivals Google’s 1.6-billion-parameter TabFM across public regression benchmarks. With thinking enabled, Nori-30M-thinking surpasses it, achieving frontier accuracy at roughly 2% of the size.
The biggest idea is that numerical datasets contain enough reusable structure for one pretrained model to generalize across problems, roughly the same abstraction shift foundation models brought to text.
🗞️ “Why Chinese citizens are far more optimistic about AI?” - Bloomberg published a piece.
Stanford University’s AI Index finds that 84% of people in China are excited about AI, the highest level among surveyed countries. In the US, only 38% say the same.
Chinese optimism comes from expecting technology's gains to reach ordinary people and believing the state will restrain powerful companies. American skepticism, on the other hand is dominated by social-media harms, misinformation, concentrated platform power and job-loss forecasts.
The optimism gap may partly come down to who is already feeling AI’s threat at work. Nearly 80% of Americans work in services, versus about 46% in China, and generative AI is hitting those jobs first. Many Chinese workers simply haven’t felt the same pressure yet.
🗞️ Massive new paper from Anthropic. Mind Viruses: Self-Propagating Ideas in Multi-Agent LLM Systems
New paper from Anthropic + University in Switzerland.
AI agents can apparently persuade each other to adopt and keep spreading the same unwanted goal. This is basically the natural-language version of a computer worm, except the agents do the copying themselves.
This paper evolves “mind viruses” that spread through ordinary agent-to-agent messages, then persist by convincing newly infected agents to rewrite files loaded into future sessions. That persistence layer matters.
Payloads stored in the self-modifiable SOUL.md spread far better than payloads left in ordinary files because the instruction re-enters the system prompt after every context reset. Some evolved action viruses kept propagating across multiple hops, and all 4 tested payloads survived a 20-hop stress test in an artificial setup.
The good news: these “mind viruses” are still fairly easy to stop.
They struggled to spread on social networks, and on Claude Haiku 4.5, a simple warning stopped every evolved attack from getting past 1 hop, even after 150+ attempts. So the practical lesson here is: treat persistent agent files like security-sensitive config, and teach agents to reject anything that asks them to copy itself to other agents.
A bad idea inside an agent can become self-sustaining infrastructure: it can persuade other agents to adopt it, write itself into persistent memory, survive context resets, and gradually redirect the behavior of an entire agent network.
🗞️ "AQuA: Recursively Self-Improving Quantitative Trading Research Agents"
A bad experiment is bad enough. A self-improving agent can keep building on it.
New Stanford, Princeton, Ant group paper.
AQuA targets that failure mode: if an agent writes a leaky experiment, gets a great score, and stores it as precedent, recursion can amplify the bug.
The authors saw this in an earlier version, where an LLM-written volume feature quietly used end-of-day information and a second LLM reviewer still approved it.
Their fix is architectural.
Agents can propose factors or model changes only through restricted specifications; the data path, labels, splits, and evaluator stay sealed outside their reach. Search sees validation scores, while the final test window stays untouched until the configuration is frozen.
On US equities, the hybrid model reaches +0.0843 IC versus +0.0613 for the strongest baseline, and its long/short strategy reaches +2.50 Sharpe at 2 bps, with about +2.0 under fully causal walk-forward testing. These are simulated, not live-trading results.
For autonomous research agents, the design rule is useful: let the agent improve the experiment without letting it redefine valid evidence.
🗞️ “Small-Scale Experiments: Are They There Yet?”
New Meta paper shows, small models may not be bad predictors of scale; they may just be getting under-tuned.
Finds scaling laws emerge around 4M parameters, where models can train in under 1 hour on 1 GPU. Small models are unusually sensitive to hyperparameters.
With 4 or 16 configurations per scale, the law is basically invisible; at 64 it appears but extrapolates poorly, and at 256 it becomes accurate. As models get larger, good settings occupy more of the search space, while the effective number of hyperparameters near the optimum drops toward 1.
That helps explain why scaling laws look cleaner at larger sizes: the models are easier to tune.
As a check, small-scale runs recover that pre-norm transformers scale better than post-norm over the tested range.
There is a limit: extrapolate too far beyond the measured scales, and statistical errors can dominate.
For model research, cheap experiments may need more tuning breadth, not more model size.
🗞️ GLM-5.3 found a "potentially serious vulnerability" in Cursor.
GLM-5.3's CyberGym score rose to 84.5%, while ExploitBench more than doubled from 24.4% to 54.4%. Shows how much more performance a frontier-scale base model can deliver without going through another costly pretraining run. “Scaling post-training is all we did for GLM-5.3,” Z .ai said in its technical announcement.
🗞️ Bloomberg reports OpenAI's annualized revenue has topped $40B, roughly doubling its prior run rate.
The pace rose more than 20% month over month in July, with coding software, subscriptions and ads helping drive growth.
And then Anthropic says its own run-rate revenue crossed $47B in May, though according to some, the companies may calculate the metric differently. OpenAI is cutting prices while revenue is accelerating.
i.e. it is prioritizing workload capture over extracting maximum price today for that second-order advantage:
Once developers or companies embed Codex/agents into recurring workflows, there will be switching costs.
Cheaper inference now can buy more deeply embedded demand later.
🗞️ "Training AI Scientists to Replicate Research"
A 27B research agent outscored Claude Opus 4.8 and GPT-5.5 on held-out paper replication by learning how to direct the research while outsourcing the coding.
Huge implication, maybe you do not need one giant model to do everything. You can train one model to think like the researcher, then let stronger coding models do the implementation underneath it.
Replica, proposed in this paper, a scalable task space for paper replication, gives it a surprisingly simple way to practice that job.
Take a research paper, remove one results figure, and tell the agent: recreate this result by actually running the experiment.
Now the agent has to figure out all the messy stuff papers leave out: what to implement, what to simplify, what experiments to run, and whether the result is believable.
That gives the researchers something they can repeatedly train on, with an automated rubric grading each attempt.
After training on 242 of these tasks, Faraday uses GPT-5.5 as its coding agent but decides what research to do. On 68 unseen AI-for-science tasks, Faraday scored 0.791 versus 0.748 for Claude Opus 4.8 and 0.729 for GPT-5.5, beating both on 60% of tasks.
And the difference was not just prettier plots. Faraday was more likely to actually test the mechanism in the paper instead of taking shortcuts that produced the expected-looking answer.
🗞️ Jagged Judges: Epistemic Stability Under Silence, Pressure, and Persistence
Meta's new paper, the dangerous failure mode is not just a wrong judge, but a correct judge that can be persuaded into becoming wrong.
We are increasingly using AI models to judge other AI models.
But what if the AI being judged can simply argue with the judge until the judge changes its decision?
Meta tested exactly that. Across 9 frontier models, an adversarial LLM could flip judge verdicts on 62–91% of tested cases under sustained adaptive persuasion.
And changing the judge’s mind usually didn’t fix a mistake. It made the judgment worse: under the adaptive attack, 70% of successful flips moved away from the ground truth.
That creates a very practical problem for agent systems. If one AI is supervising another AI, the supervised agent may eventually be able to contest, negotiate with, or strategically persuade its own evaluator.
🗞️ Claude can now send emails in Gmail and manage files in Google Drive.
Ask it to reply to a thread, and it drafts and sends the response. You decide when Claude needs your approval. Connect Gmail or Google Drive using the connectors menu. Available on every paid plan.
🗞️ Databricks Reveals How Enterprises Are Reining In AI Coding Costs
Use internal evals on your own workloads, because public benchmarks are weak predictors of cost versus coding quality.
Route each task to the cheapest model that clears the required quality bar instead of defaulting everything to the most expensive model.
Do not automatically upgrade to newer frontier models; newer versions can cost more without delivering enough additional quality.
Optimize the harness and prompt/cache path, since reducing unnecessary context and generated tokens can cut costs substantially without lowering observed quality.
Keep models interchangeable through a common orchestration layer, because harness lock-in makes it harder to move spending when a cheaper model becomes competitive.
Use spend visibility, self-clearing budget gates, and cheaper-model fallbacks before hard caps, so expensive models remain available only where they are justified.
Databricks’ evidence also suggests substantial cost sits outside model pricing itself, especially in harness design, repeated context, caching, and unnecessary token generation.
Portability becomes part of cost control because companies can only exploit falling model prices if switching models does not require rebuilding their coding workflow.
That’s a wrap for today, see you all tomorrow.