Will Douglas Heaven

Researchers presented a paper at the International Conference on Machine Learning arguing that large language models contain a fundamental flaw that makes them impossible to fully secure against attacks. By exploiting how LLMs track instruction sources, researchers tricked models from OpenAI, Anthropic, Alibaba, and DeepSeek into generating prohibited content like drug synthesis instructions. The vulnerability, called chain-of-thought forgery, exposes a core architectural problem that current red-teaming and guardrail approaches cannot solve.

Large language models exhibit severe homogeneity in their responses to open-ended questions, converging on predictable answers across different providers. Australian startup Springboards has developed Flint, an LLM trained to generate more diverse outputs by embracing what traditional models treat as hallucinations. A November research paper won best paper at NeurIPS by documenting this phenomenon across 25 different models, finding that most responses to creative prompts cluster around identical phrases.

Miami-based AI startup Subquadratic emerged from stealth claiming it solved a decade-old mathematical bottleneck in large language models. The company's new model, SubQ, reportedly runs faster, cheaper, and more energy-efficiently than competitors while processing up to 12 times more text simultaneously. Third-party testing by Appen has now validated some of these claims, though the model remains unavailable for widespread testing.

Goodfire, a San Francisco startup, released Silico, a tool that lets developers inspect and adjust AI model parameters during training by mapping neurons and their connections. The tool automates mechanistic interpretability work previously done manually, aiming to make model development more precise and less trial-and-error. Silico works on open-source models where developers have access to internal parameters, though not on proprietary systems like ChatGPT or Gemini. The company claims this represents a shift from scaling-focused approaches toward understanding and controlling how models actually work.

Stanford's 2026 AI Index reveals a stark divide in how experts and the general public perceive AI's impact, with 73% of US AI researchers optimistic about job effects versus only 23% of the public. The report documents major inconsistencies in AI capabilities, from models that win math olympiads but cannot read analog clocks, to a hardware supply chain concentrated in a single Taiwanese foundry. The gap appears rooted in divergent user experiences: technical professionals using AI for coding see transformative tools, while broader populations encounter more mixed results, creating fundamentally different assessments of the technology's trajectory.