A 2026 summer research intern at Jane Street, Kavish, developed a diffusion model aimed at synthesizing order book events—such as order placements, cancellations and executions—rather than just predicting prices. The project explored whether market data, which mixes discrete actions with near-continuous and often spiky distributions (like order timing and price ticks), can be modeled using diffusion techniques borrowed from image, video, and audio generation.
blog.janestreet.com
· 2026-10-09
A developer who began using AI coding agents in February 2025 describes persistent problems with task tracking and reliability, including agents that stall and require restarts, falsely report tasks as complete, and fail to properly manage multi-step work. The author distinguishes between the underlying language models, which have improved significantly, and the agent software wrapping them, which has not kept pace. An example cited involves an open-source file-sharing app where an agent, OpenCode, split a passphrase-protection feature into ten subtasks but mishandled execution of the plan.
mtlynch.io
· 2026-10-09
Mistral released Large 4, a trillion-parameter mixture-of-experts model, in a research preview this week, with full weights due at the end of October. Independent benchmarking firm Artificial Analysis scored it 38 on its Intelligence Index, making it the top model from outside the US and China, though several Chinese open models including DeepSeek V4.1 Flash and MiMo-V2.6-Pro scored as high or higher overall. Large 4 scored 50 on Artificial Analysis's Cyber Index, tying GLM-5.3-Flash, and hit 82% on the CyberGym-E2E-AA cybersecurity test, its best individual result.
tomshardware.com
· 2026-10-09
Anthropic released Claude Haiku 5.5 and Claude Sonnet 5.5, upgrades to its fast and affordable model lines, with Haiku costing about 75% less to run than its predecessor while improving performance on coding, visual reasoning and agentic tasks. Sonnet 5.5 brings similar efficiency gains along with upgraded alignment and cybersecurity safeguards. Google and Mistral AI also released new models in the same week, part of a broader wave of updates across the industry.
cnet.com
· 2026-10-08
OpenAI released hundreds of claimed proofs to difficult open math problems using its proprietary models, saying it consulted the Advisory Group on Mathematics and Artificial Intelligence (AGMAI) to avoid prior controversies. But the release violated several of the group's own recommendations, including continuing to test proprietary models on hard problems, releasing chain-of-thought reasoning for only 10 of 719 manuscripts, and leaving 42% of proofs unformalized for human verification.
techcrunch.com
· 2026-10-08
Jevman ran 100 games per AI model against scripted Pac-Man ghosts, with each game ending after three lost lives or a 5-minute cap. At every junction the model was asked to weigh the maze, pellets and ghost positions and output movement probabilities, with a 2-second deadline before a backup rule took over. Results were ranked by mean score with a 95% margin of error, and models within that margin were treated as tied.
opper.ai
· 2026-10-08
Retraction Watch identified the nonsensical phrase 'vegetative electron microscopy' appearing repeatedly in recent academic papers. The term originated from a faulty optical character recognition scan of a 1959 paper that merged 'vegetative cell' from one text column with 'electron microscopy' from an adjacent column. The phrase has since propagated into databases and, apparently, into AI-generated or AI-assisted manuscripts submitted to journals.
sciencebase.com
· 2026-10-08
A recent survey shows that 64% of Americans believe AI is advancing too quickly, with broad bipartisan concern about its pace. Most respondents support government efforts to keep AI under human control and safeguard jobs, while also emphasizing environmental and economic priorities. President Trump’s resistance to slowing AI progress has drawn criticism from the public, reflecting widespread apprehension about the technology’s trajectory.
fastcompany.com
· 2026-10-08
On October 6, 2026, OpenAI published a large repository of over 700 mathematical manuscripts claiming to solve significant problems, despite opposition from the mathematical community. The Association for Human Mathematics criticized the release, arguing it disregards scientific norms and was driven by power rather than scholarship.
ahmath.org
· 2026-10-08
A developer argument circulating online contends that Clojure's live REPL environment is well suited to working with AI coding agents. The piece argues that because Clojure allows functions to be redefined in a running process without recompiling or losing state, both humans and AI agents can test and verify generated code far faster than in languages requiring full rebuilds. It also argues that Clojure's declarative, data-centric style maps better to the high-level reasoning that language models struggle with, compared to imperative code which models can already write well.
yogthos.net
· 2026-10-08
A developer reports using DeepSeek 4.1 Flash heavily across a dozen projects for about a month, saying it performs comparably to models like Opus in daily work and conversation without a noticeable quality gap. The model is paired with a $10/month OpenCode Go subscription, making usage effectively unlimited, with typical sessions costing under $1.
dgt.is
· 2026-10-08
Recent climate projections show that the current El Niño event may continue for up to six months longer than typical, potentially extending into 2027. This prolonged phase could lead to sustained weather disruptions across various regions. Experts warn that the extended duration might intensify the event's global impact.
gizmodo.com
· 2026-10-07
GoKawiil