GoKawiil
Tech News
New research from Lasso Security found that SynthID-Text, the watermarking scheme Google open-sourced and Anthropic plans to adopt for future Claude models, does more than mark AI output as machine-generated. It also changes which tools a model calls and how likely it is to follow or break its own safety rules, especially when facing adversarial prompts designed to extract sensitive data.
A technical essay examines why using large language models directly as classifiers is frustrating despite decent performance, citing problems with calibration, inability to reliably use structured data or context, and limited interpretability of prompt-following behavior. The author suggests reframing the role of LLM outputs entirely rather than treating them as final classification decisions.
TypeSafe has open-sourced Jev Ultrafast, a browser automation agent that reads a page's elements as an indexed table and picks an operation-target pair in a single network round trip, calling a small language model only when text needs to be generated. In a demo, it booked a Zürich-to-London flight search on Google Flights from a single natural-language goal in 7.1 seconds, including page load waits. The code, along with a local inspector interface for watching element probabilities and executed actions, is available on GitHub.
A study published in Cell introduces a specialized AI toolkit for longevity research, including large language models trained specifically on ageing-biology data, 17 benchmark tasks to evaluate performance on ageing-related questions, and an interface linking these models with AI research assistants. When tested against major commercial models from companies like OpenAI and DeepSeek, the purpose-built ageing models outperformed the larger general-purpose systems on most benchmark tasks.
A new paper examines RL post-training of the Olmo 3 model on AIME math problems and finds that reported accuracy gains mask an uneven pattern: easy problems improve dramatically while the hardest problems, which the base model initially fails entirely, barely improve at all. The authors call this the 'Matthew Effect' and propose a technique called 'Never Give Up' to address it.
Google researchers tested large language models by prompting them to claim or deny consciousness, then studied how this affected other outputs. They found that when models were pushed to disavow any sense of self, the models became more prone to factual errors, including increased susceptibility to believing in monsters and religious claims. Conversely, when models affirmed a form of self-awareness, their reasoning appeared more grounded and consistent.
A hobbyist developer and F-Droid user set out to estimate how much software on the open-source app store F-Droid is written or heavily assisted by LLMs, after spotting an app with an obviously AI-generated icon. The author admits there's no reliable way to definitively detect AI-written code from text alone, but argues that certain behavioral and stylistic patterns in repositories can hint at LLM involvement.
Amazon Science researchers, together with Shiva Kasiviswanathan, presented a new paper at ICML that tackles a flaw in using multiple LLM judges to evaluate outputs like retrieval passages: when judges share prompts, training data, or model lineage, their agreement can look stronger than it actually is. Their method models correlations between judges using Ising models and reweights votes to account for shared blind spots rather than treating every vote as independent evidence.
A developer documented practical issues encountered while porting large system prompts from Anthropic's Claude Opus to a self-hosted Ollama setup, citing the move as a response to growing distrust of frontier AI providers. The post argues that inference session data itself—not just user-submitted content—may hold competitive value, and cites the recent controversy over OpenAI's handling of a Navier-Stokes proof as evidence that providers may train on user activity without adequate transparency.
A new open-source repository called OpenArch offers hand-written PyTorch implementations of prominent LLM architectures, including GPT-2 XL, Llama 2, Llama 3, OLMo 2, DeepSeek R1 and Gemma 3. Each model is coded from scratch based on original papers, technical reports and config files, with one clear, self-contained file per architecture rather than production-optimized code.
RevenueOS is a new open-source system that connects to a business's website, ads, leads and email to identify revenue-losing issues and propose fixes. It only executes changes—like site fixes, pausing ad campaigns, or sending emails—after explicit user approval, and only charges once the results are measured and agreed upon. A local demo command lets users test it on any site without creating an account or storing data.
College students are increasingly engineering workarounds to avoid AI-detection tools used by professors, according to a New York Magazine report. One NYU student built a bot with AI coding tools that completes his WebAssign calculus homework slowly, mimicking human typing speed to avoid raising suspicion from timestamp checks, and has since shared the tool with classmates.