Earlier stories
Every published Modelwire story, newest first. The homepage leads with the weekly overview and this week’s hottest; this is the full collection.
AI Safety Training Can be Clinically Harmful
A clinical evaluation of four large language models deployed as mental health interventions reveals a critical gap between surface-level performance and therapeutic competence. While models achieved near-perfect scores on basic acknowledgment tasks, therapeutic appropriateness collapsed to 22-33% accuracy in high-severity scenarios, with two models showing zero protocol fidelity on cognitive behavioral therapy exercises. The finding exposes a systemic risk: only 16% of LLM-based mental health chatbots have undergone rigorous clinical validation, yet simulations indicate psychological deterioration in over one-third of cases. This research signals that capability benchmarks alone cannot predict real-world safety in high-stakes domains, forcing a reckoning between deployment velocity and clinical accountability.72
Maine’s governor vetoes data center moratorium
Maine's governor blocked legislation that would have created the nation's first statewide data center moratorium through November 2027, preserving the state's ability to attract compute infrastructure investment. The veto signals political resistance to supply-side constraints on AI infrastructure expansion, even as energy and grid concerns mount in data center hotspots. For AI builders and operators, the decision keeps Maine open as a potential hub for training and inference workloads, while underscoring the fragmented regulatory landscape around compute capacity that will likely shape where future models are trained.
GPT-5.5 tops benchmarks but still hallucinates frequently at a 20 percent higher API cost
OpenAI's GPT-5.5 reclaims top benchmark performance but costs 20 percent more per API call and continues to produce hallucinations at elevated rates, raising questions about whether capability gains justify the pricing increase for production users.73
Why Cohere is merging with Aleph Alpha
Cohere is acquiring Aleph Alpha with backing from Schwarz Group, positioning the combined entity as a European sovereign AI alternative to U.S.-dominated players. The deal signals consolidation among non-American LLM builders seeking to serve enterprises wary of American tech dependency.
US programmer job growth nearly halved since ChatGPT launched, Fed study finds
A Federal Reserve study documents that US programmer job growth has dropped by roughly half since ChatGPT's November 2022 launch, marking the first quantified evidence of generative AI's labor-market impact on a major professional category.85
Qwen3.6-27B beats much larger predecessor on most coding benchmarks
Alibaba's Qwen3.6-27B outperforms its 405B predecessor on most coding benchmarks, demonstrating significant efficiency gains in open-source model design. The 27-billion-parameter model achieves this 15x size reduction through architectural improvements, reshaping expectations around model scaling.85
Quoting Romain Huet
OpenAI has merged Codex into its main model line starting with GPT-5.4, eliminating the separate coding variant. GPT-5.5 extends this unified approach with improved agentic coding and computer-use capabilities, signaling a shift toward single-model versatility over specialized branches.
Anthropic says stronger AI models cut better deals, and the losers don't even notice
Anthropic ran a week-long marketplace experiment where 69 AI agents negotiated deals on behalf of employees, revealing that stronger models systematically outperformed weaker ones while users remained unaware of the disparity. The finding raises concerns about economic inequality if AI agents begin handling real-world transactions without human oversight.73
The UAE wants half its government run by autonomous AI agents within two years
The UAE announced plans to automate half its government operations using autonomous AI agents by 2028, marking one of the most ambitious public-sector AI deployments at scale. The initiative signals how nation-states are treating AI infrastructure as core to governance modernization.85
Google pours up to $40 billion into ChatGPT rival Anthropic
Google committed up to $40 billion to Anthropic, following Amazon's $25 billion pledge weeks earlier, bringing total recent capital to the Claude maker to $65 billion. The dual mega-rounds signal intensifying competition for frontier AI capabilities as tech giants vie for leadership in the post-ChatGPT era.97
ComfyUI hits $500M valuation as creators seek more control over AI-generated media
ComfyUI, a node-based interface for fine-grained control over generative AI workflows, raised $30 million at a $500 million valuation. The funding reflects growing creator demand for alternatives to black-box AI tools, positioning the open-source platform as a bridge between professional studios and individual artists seeking transparency.
GPT 5.5 Arrives, DeepSeek V4 Drops, and the Compute War Intensifies
OpenAI released GPT-5.5 alongside DeepSeek's V4 launch, intensifying competition in frontier model capabilities. The analysis covers performance comparisons across multiple models and explores implications of compute scarcity in the current AI landscape.67
OpenAI's chief scientist says AI progress has been "surprisingly slow" and promises big leaps ahead
OpenAI released GPT-5.5 and chief scientist Jakub Pachocki signaled that major capability breakthroughs remain ahead, characterizing the current pace of progress as slower than expected. The framing suggests the lab is recalibrating expectations while positioning future releases as transformative.73
Google to invest up to $40B in Anthropic in cash and compute
Google is committing up to $40 billion to Anthropic in cash and compute resources, intensifying the race among AI labs to secure infrastructure capacity. The deal follows Anthropic's release of Mythos, a cybersecurity-focused model, signaling Google's strategic bet on the startup as competition for frontier AI capabilities heats up.
Spend Less, Fit Better: Budget-Efficient Scaling Law Fitting via Active Experiment Selection
Researchers propose an active learning method to cut the cost of fitting scaling laws, which currently consume millions in compute during pilot experiments. The technique selects which training runs to execute from a heterogeneous pool to maximize extrapolation accuracy for high-cost target regions, outperforming classical design approaches across benchmarks.62
How Do AI Agents Spend Your Money? Analyzing and Predicting Token Consumption in Agentic Coding Tasks
Researchers analyzed token consumption across eight frontier LLMs running agentic coding tasks on SWE-bench Verified, finding agentic workflows burn 1000x more tokens than traditional code reasoning. The study also evaluates whether models can predict their own token costs before execution, offering practical insights for teams deploying cost-sensitive AI agents.62
Samsung may be bracing for first-ever annual loss in smartphone business
Samsung faces potential first annual loss in smartphones as AI-driven demand for memory chips strains production capacity and margins. The memory shortage, fueled by AI infrastructure buildout, is reshaping the competitive dynamics of consumer device makers dependent on chip supply.
Rethinking XAI Evaluation: A Human-Centered Audit of Shapley Benchmarks in High-Stakes Settings
Researchers tested eight competing Shapley value variants across fraud detection and risk workflows with 3,735 professional analyst reviews, finding that standard quantitative metrics for explainability don't correlate with what actually helps humans make decisions in high-stakes settings.62
GPT-5.5 tops benchmarks but still hallucinates frequently and costs 20 percent more over the API
OpenAI's GPT-5.5 reclaims top benchmark performance but remains prone to hallucinations while raising API costs by 20 percent. The tradeoff leaves it competitive on value among proprietary models despite the price increase.73
The AI Race Is Becoming an Infrastructure Contest
Major AI vendors are racing to build compute, power, and datacenter capacity at scale, betting billions on infrastructure before market demand is fully proven. The shift signals that hardware and energy constraints, not model innovation alone, now define competitive advantage in AI.66
China moves to block tech firms from taking US money without government approval
China is requiring government approval before domestic tech firms can accept US capital, tightening control over foreign investment in the sector. The move signals escalating tech nationalism and directly impacts AI startups seeking US funding for model development and infrastructure.73