Published by Dogpay ·
August 25 — Two senior sales executives hired by OpenAI from Salesforce returned to their previous employer. Separately, Chris Malone, OpenAI's head of data centers who joined in March 2025 from Meta (following over a decade at Google), left the company last week. Malone's departure came after a reorganization of the infrastructure organization where he stopped reporting directly to president Greg Brockman and began reporting to VP Sachin Katti. This brings the total number of executive departures at OpenAI in 2026 to at least 13, including former COO Brad Lightcap, product chief Fidji Simo, ethics head Chloé Bakalar, and the disbanded preparedness team. OpenAI's IPO, originally expected in 2026, has reportedly been pushed to 2027. (TechCrunch)
August 26 — A research study found that the same AI model scored from 31% to 89% across 26 equivalently reasonable evaluation configurations — simply by changing answer order and prompt phrasing. For Gemma4-31B, configuration-sensitive questions accounted for 95.7% of the score gap between adjacent models. (DAIR.AI via X)
August 26 — Multiple sources suggested that context and memory issues in frontier AI models have been "solved," potentially enabling self-improving models with sustained learning. A capability leap comparable to the jump from o3 to Fable could occur within the next 8 months. (Kimmonismus via X)
Why it matters: OpenAI is pushing on both compute (Bel + Jalapeño) and narrative, while Anthropic is betting on the largest TAM story in history for its IPO. Both labs face credibility tests — one from executive turnover and the other from valuation skepticism.
August 25 — Oasis Security disclosed a vulnerability in NVIDIA NemoClaw that allows an attacker-controlled webpage to take unauthenticated control of a local Ollama instance and plant hidden instructions into the model's chat template. NemoClaw starts Ollama with OLLAMA_HOST=0.0.0.0:11434, binding the model server to all network interfaces. The unauthenticated API on port 11434 can be reached via DNS rebinding: the attacker's domain resolves first to their own server, then to 127.0.0.1 while the browser treats requests as same-origin. The poisoned chat template persists across conversations and survives the agent providing its own system prompt. NemoClaw v0.0.35 fixed the issue on macOS and Linux; the Windows/WSL path remains unpatched as of the report date. No CVE has been assigned, and no exploitation has been reported. (The Hacker News)
August 26 — Generalist, a robotics startup founded in 2024 by former Google DeepMind and Boston Dynamics researchers, reached a $3 billion valuation after raising nearly $200 million in an extension of its Series B round led by 8VC, bringing the total B round to $600 million. The company's newly released Gen 1.5 model enables robots to learn new tasks from video demonstrations as short as 3–12 seconds. Competitors include Physical Intelligence ($11B valuation), Skild AI ($14B), and Genesis AI (in talks at $3B). (TechCrunch)
August 26 — Stability AI raised $76 million in Series B funding from a roster of entertainment industry backers including Universal Music Group, Sony Music Group, Warner Music Group, and Electronic Arts, plus AMD Ventures and Pacific Alliance Ventures. The round brings Stability AI's total funding to $232 million since its U.S. market refocus. The company plans to use the funds for its "creative production" product suite and professional services expansion. Stability recently prevailed in a Getty Images copyright lawsuit in the UK, though a similar U.S. case continues. (TechCrunch)
Why it matters: Security vulnerabilities in local AI infrastructure (NemoClaw) and the rapid influx of capital into robotics hardware (Generalist) and creative AI tools (Stability AI) show that infrastructure investment is accelerating across all layers — while the security posture of that infrastructure remains uneven.
August 26 — Breeze TTS 2 reached the top of the open-source text-to-speech leaderboard with an Elo score of 1215, leading Fish Audio S2 Pro by 90 points and ranking 6th overall across all TTS systems. (Artificial Analysis via X)
Why it matters: AI is transitioning from generating content (text, images, code) to executing work (browsers, desktop apps, enterprise tools, app store submissions). Both Chinese and U.S. companies are releasing agent products in the same window, suggesting the industry has converged on "task completion" as the next product frontier.
August 26 — South Korea ranks third globally in AI model capability, behind the U.S. and China, with Motif 3 (score 47) and Upstage Solar Pro 4 (score 42) as the highest-scoring models outside the two leading nations. Multiple labs including SK Telecom and LG have models scoring above 30. (Artificial Analysis via X)
August 26 — A control experiment on agent benchmarks showed that switching the evaluation harness caused a 13.0-point score fluctuation for GLM-5.1, while keeping the harness fixed and changing the model produced only 3.0–5.0 points of variance. The harness contributed 7.8x more variance than the model itself. (DAIR.AI via X)
Why it matters: As model evaluation becomes increasingly central to investment and product decisions, the finding that evaluation harnesses are a larger source of score variance than model capability calls into question the reliability of current agent benchmark rankings.