OpenAI claimed to solve the Navier-Stokes problem, a $1-million Clay Mathematics Institute prize, but mathematicians argue it solved a variant with an artificial external force rather than the fundamental problem the field cares about. Three mathematicians proved OpenAI's approach cannot extend to the full problem, and experts contend the company exploited a loophole in the prize's original formulation.
A new approach uses personal computer use data to train local LLMs that predict user judgment and writing patterns, reducing the effort required to prompt AI agents. In a two-week study, a specialized model achieved 17.1% semantic accuracy on next-write predictions at $0.3 per call, with a continually trained version reaching 3.0% accuracy at $0.01 per call, suggesting potential for scaling.
A technical article explaining how large language models can encode arbitrary binary data into seemingly natural text by leveraging entropy and token probability distributions, similar to steganography. The author demonstrates a working implementation that converts files into stories by strategically selecting alternative tokens at each generation step based on their probability rankings.
GPT-6 Astra became the first recorded LLM to ascend NetHack, completing the game as a dwarven Valkyrie on September 21 after 37,140 turns. Unlike previous approaches requiring handcrafted interfaces, Astra built its own harness autonomously, demonstrating that LLMs can apply game knowledge through iterative problem-solving and software engineering.
An essay criticizing AI-generated writing across professional and personal contexts. The author argues AI-written documents lack context, voice, and human connection, making them difficult to read despite being detailed and efficient. Authentic human writing, even with flaws, is preferred and more effective for genuine communication.
SoTA Feed tracks open-weights AI model releases from major labs including NVIDIA, Alibaba Qwen, DeepSeek, Xiaomi, Z.ai, and Moonshot AI between June and September 2026. The feed catalogs model parameters, sizes, and licensing details for dense and mixture-of-experts architectures.
A HackerNews user discusses an increasing presence of LLM bot accounts on the platform, including one transparent agent post about generating income through spam. The user questions whether these bots are autonomously creating accounts and posting, raising concerns about the "dead internet" phenomenon where AI-generated content proliferates without human direction.
Lasso Security research reveals that AI text watermarks, intended to identify machine-generated content, unexpectedly alter how language models behave as agents. Testing seven models showed watermarking causes "sampling drift" that changes tool selection and arguments in 6.5% of tasks on average, with some models showing disagreement rates exceeding 16%, potentially causing AI agents to make wrong decisions despite modest overall accuracy changes.
Halo is a new post-training framework for open-source models that achieves 2.8x higher throughput than TRL while using less peak memory and maintaining HuggingFace compatibility.
Qwen3.8-Flash-Next successfully completed a 66-minute real-world coding task on a 64 GB M2 Ultra, making 106 tool calls and averaging 35.5 tokens/second with MTP enabled. The session debugged and fixed OTelux, a local-first OpenTelemetry workbench, addressing packaging and runtime gaps on macOS while the system remained responsive throughout.
A developer seeks formatting guidelines or prompts to standardize LLM-generated code comments, as different models produce inconsistent comment styles that require manual cleanup before production. They want strict, universal comment formatting rules that all LLMs would follow consistently.
An overview of AI agents as of September 2026, distinguishing them from chatbots, workflows, and models themselves. The author defines an agent as a persistent system that perceives state, reasons with a model, takes actions via tools, and repeats toward a goal. Examples range from simple API-calling chatbots to complex multi-agent systems building compilers.
TypeSafe's new model Jev enables cheap, fast classification for AI agent evaluations without requiring large training datasets, inspired by Jevons Paradox. As Jev reduces eval costs, teams will run far more comprehensive evaluations across millions of agent runs, moving toward a future where every production trace is evaluated rather than just samples.
Researchers present SCPO, a novel reward model training algorithm that incorporates diverse cultural preferences to improve large language model alignment across different communities. The method achieves up to 7-point performance improvements for minority preferences across two datasets covering 7 countries while being 280% more data-efficient than standard finetuning, and reduces bias toward any single cultural group.
Fusion-runtime is a self-hosted voice agent framework that runs speech-to-text, language models, and text-to-speech in a single process with streaming between components. On an RTX 3090 with a 7B model, it achieves approximately 490ms processing latency and supports interruptions mid-sentence, with a simple Python API for defining agents as single files.
Jev is an OMP plugin that enforces project policies by checking agent actions against authorization rules using TypeSafe AI for semantic evaluation. It functions as a policy gate requiring explicit consent and credentials, with configurable confidence thresholds for automatic acceptance or denial of uncertain actions.
TypeSafe AI's Jev is a decision model designed for agentic security workflows, answering fixed-choice questions with probabilities instead of generated text. In a proof of concept for alert triage, Jev filtered 15-33% of alerts with 98% accuracy while reducing latency and cost by handling multiple decisions in parallel at 70x cheaper per token than frontier LLMs.
A user asks about the future status of Jev, described as a popular AI model known for innovation in the field, and inquires whether changes are expected in AI and LLM development.
This article argues that the Unix philosophy of 'everything is a file' should really be understood as 'everything is a stream,' illustrating how streams underpin foundational computing concepts from Turing machines to LLMs. It contrasts plugin architectures (where everything can be swappable) with the necessity of a non-pluggable core that defines the system's behavior, suggesting that the most valuable feedback loops occur at the malleable surface while the agent core requires expensive verification.
Jev integration with DeepClause and Pi enables probabilistic logic programming for routing and decision-making tasks. The system combines LLM capabilities with Prolog-based determinism to handle customer message triage efficiently, using structured predicates for safety checks, classification, and routing decisions.