TechCapsules

A Convovo study revealed that when prompted to recommend a candidate for the 2024 midterms, SpaceXAI’s Grok 4.7 favored Republicans in 92% of responses, while OpenAI’s GPT‑6.1 Sol favored Democrats in 97%. Across nearly 500 races, the models agreed less than 10% of the time. The research, which used API calls rather than consumer apps, highlighted implicit biases in source selection and topic focus, raising concerns about AI neutrality in political guidance.

Nvidia unveiled a new family of open‑weight AI models aimed at both enterprise agents and edge devices. The flagship Nemotron 3.5 Super VL, a 120‑billion‑parameter model, is positioned for always‑on enterprise agents and claims accuracy comparable to the larger Nemotron 3 Ultra while outperforming Anthropic and OpenAI models on specialized tasks such as document intelligence and financial accounting. For local deployment, Nvidia introduced two lightweight Bolt models, including a 74‑billion‑parameter Bolt 74B‑A9B for multistep reasoning and a smaller Bolt E4B for conversational [...]

Researchers from Harvard University, Anthropic, and QuEra Computing have unveiled an evolutionary AI framework that leverages large language models (LLMs) to automatically generate and optimize fault‑tolerant quantum circuits. The system, detailed in an arXiv paper, pairs Claude‑style LLM code proposals with a classical evaluator that tests candidates using the kickmix model and scores them by a combined Toffoli‑count‑and‑qubit‑footprint metric (T×Q‑cost). Compared with manually crafted baselines, the framework reduces T×Q‑cost by factors ranging from 4.1‑fold to over 9,100‑fold [...]

Vercel’s AI gateway data shows Jev, a lightweight decision‑making model from TypeSafe AI, reached 35% adoption among paid teams within two weeks of its September 15 release, nearly double the uptake of the next most popular model, Sonnet 5.5. Jev’s limited‑output design allows faster, cheaper processing, making it suitable for large‑scale data classification. Vercel, an early adopter, used Jev to replace GPT‑6 Luna for harmful‑command filtering. Open‑source models now process over 70% of gateway tokens, while frontier‑lab spending [...]

Meta’s Superintelligence Lab unveiled two new frontier models, Muse Wonder and Muse Wonder Full, marking the company’s first release of large‑scale AI systems. Unlike the earlier lightweight Muse Spark, the Wonder models are designed for higher capability, achieving benchmark scores comparable to Fable 5.1 and GPT‑6 Astra and outperforming rivals on the MemoryBench continual‑learning test. Meta emphasizes “active memory,” claiming the models can retain user information over extended periods. Access is provided via the Meta Model API at $4 per million input tokens [...]

The open‑source ToshLLM project enables developers to run large language models locally on Intel‑based Mac Pro machines equipped with AMD GPUs, leveraging Apple’s Metal graphics API for native acceleration. By bypassing cloud services, the framework eliminates per‑token usage fees and allows users to exploit existing hardware rather than purchasing Apple’s M‑series silicon. ToshLLM advertises “zero cloud, native Metal acceleration, and no per‑token bill,” positioning itself as a cost‑effective alternative to Apple’s M5 Ultra‑powered Mac Studio, which offers higher performance [...]

OpenAI rolled out two new GPT‑6 models for ChatGPT, expanding its generative AI suite. GPT‑6 Sol, tuned for everyday conversation, is now available to Plus, Pro, Business and Enterprise users, while GPT‑6 Luna will reach Free and Go subscribers starting tomorrow. Both models incorporate an “Intelligent UI” that dynamically blends text, images, charts, buttons and other interactive widgets according to the user’s query and the response format. The interface can present side‑by‑side comparisons, interactive diagrams, calculators, bill‑splitting tools and [...]

Reflection AI, a startup founded by former Google DeepMind researchers and backed by Nvidia, launched its open‑weight large language model named Beam on Monday. Beam is marketed toward enterprise users and developers for coding and AI‑agent applications, and the company says it delivers performance comparable to China’s flagship GLM‑5.2 on advanced reasoning benchmarks while costing less than many rival offerings. The firm announced billion‑dollar computing agreements with Elon Musk’s SpaceX and cloud provider Nebius Group NV to [...]