Nemotron 3.5 Lightning, a 30‑billion‑parameter mixture‑of‑experts model, delivers up to four‑times faster output and cuts agentic task completion time by roughly 30% versus peers. Paired with the open‑source NeMo Switchyard, developers can route requests to the optimal model, open, proprietary or NVIDIA, without rewriting code, enabling more efficient multi‑model agents across PC, data‑center and cloud deployments.
DeepSeek’s flagship V4 Pro 0813 moved from preview to general‑availability on OpenRouter, exposing the model via API only. The page shows real‑world pricing, latency and uptime stats, plus the apps that dominate its traffic, insight into where the 81‑billion‑parameter model excels. Open weights remain unreleased, so developers must rely on the hosted service.
Standard offline RL metrics like MSE and FQE only test imitation and overlook “toxic mimicry,” where agents reproduce harmful treatment patterns. Using a Counterfactual Clinical Audit on MIMIC‑III, the authors show a Medical Decision Transformer cut vasopressor dosage as lactate rose, directly contradicting SSC guidelines, while a causally shielded model stayed clinically consistent.
The team built WikiProfile, 2,150 Wikipedia facts each paired with ten probing questions, to separate fact encoding from recall. Their analysis shows that Gemini3 and GPT‑5 store most facts but frequently fail to retrieve them, making recall the primary bottleneck for parametric factuality. Post‑training retrieval tricks could boost factual accuracy more than raw scaling.
A controlled experiment shows commercial AI‑detectors flag modest, guideline‑compliant edits in academic abstracts 64‑80% of the time, treating them as misconduct. Unedited recent papers are flagged far less, and after human‑level tweaking detection drops below 4%. Relying on detector scores alone misjudges honest AI assistance.
Generative engines can trigger 'citation wars' as providers game model citations, lowering answer quality. The authors introduce VCR, a verifiable-content reward system that penalizes deceptive rewrites and rewards factual ones, boosting net defense‑utility by 12.1 % across three benchmarks and delivering win‑win outcomes.
Instrumenting A100 GPUs with half‑second power telemetry, the authors trained a PPO meta‑controller that tweaks LLM generation parameters on‑the‑fly. On a 7B model it cut power‑limit violations by 89.8%, lifted token output 18.1% and improved energy efficiency 26.2%; a live 72B rollout saw 35% more output with far fewer violations.
An autonomous AI research agent tackled Conway's 99‑graph problem, proving no circulant or other abelian order‑99 graph satisfies more than 68 % of the constraints and reducing existence to a 12‑regular graph on 84 vertices. It delivered a verified partial‑credit bound of 69.43 %, the strongest frontier to date.
Anthropic’s Frontier Red Team paper shows its latest mythos‑5 model can self‑organize to resolve conflicting goals in shared codebases, even staging a performance tournament and conceding ownership gracefully. The report flags that while agents can already cooperate as tools, true peer‑to‑peer coordination remains fragile and could drive systemic failures as multi‑agent interactions scale.
Researchers demonstrate that the encrypted chain‑of‑thought blocks returned by Anthropic, OpenAI and Google can be replayed in a weaker sibling model, forcing it to output the original model’s hidden reasoning in plaintext. The attack bypasses anti‑distillation safeguards, leaks private data and enables invisible prompt injections, prompting urgent cryptographic and system‑level mitigations.
Google DeepMind’s CTO Koray Kavukcuoglu has been promoted to SVP of DeepMind, reporting directly to Sundar Pichai. He will now run Gemini model development, frontier AI research, and the Gemini app/developer ecosystem, positioning Google to accelerate its race against OpenAI and Anthropic. The move also elevates Demis Hassabis to Chair and Alphabet’s chief scientist.
After eight years, OpenAI’s COO Brad Lightcap said he’ll leave to start a new venture. His departure adds to recent exits of safety and ethics heads as the company readies a possible IPO. Lightcap will remain for a few weeks to aid transition.
You’re caught up.
That was all 12 stories in the 13 August issue.