Benefit, harm, and capability ledgers
Events
Frequency · stacked by domain
Events over time
- Computer Science
Heatmap
Domain by month
Significance
Impact tiers
Base impact
- T1 · 1Useful / Limited
- T2 · 3Notable / Significant
- T3 · 10Major
- T4 · 30Historic
- T5 · 100Civilizational / Catastrophic
Base points before attribution, evidence, realization, durability, and credit share.
MethodologyFrequency
Benefit versus harm counts
2D semiconductor aims to turbocharge development of high-performance electronics
2D semiconductor aims to turbocharge development of high-performance electronics
Fine-Tuning Fixes Mode Collapse and Over-Dispersion in LLMs
arXiv:2609.16454v1 Announce Type: new Abstract: Recent work by Doshi and Hauser (2024), Bisbee et al. (2024), and Xie et al. (2026) raises concerns that outputs from large language models (LLMs) tend to be under-diverse: they repeat or resemble one another more often than responses from the population they are meant to represent, a phenomenon known as mode collapse. In this work, we show that whether mode-collapse, or its opposite, occurs depends on the specific model and dataset used. Further, with sufficient supervised fine-tuning (SFT) data, LLM output diversity converges toward that of th
Where Should the KV Cache Live? Placement Policies Across GPU, CPU, and SSD for Long-Lived Sessions
arXiv:2609.16215v1 Announce Type: new Abstract: GPU high bandwidth memory is scarce and expensive, and KV caches consume much of it as chats, agent loops, and document question answering accumulate state. Systems such as Mooncake, LMCache, FlexGen, InfiniGen, and AttentionStore extend GPU memory with CPU DRAM and SSD. The harder question is which blocks belong in each tier, when to move or evict them, and whether prefetching helps. We study these choices in a discrete event simulator spanning GPU HBM, CPU DRAM, and SSD, calibrated against a random forest execution time predictor. We compare r
Rsearch: An R interface to VSEARCH supporting visualization and parameter tuning
Background: We present Rsearch, an R package that integrates the core functionality of VSEARCH into the R environment and extends it with visualization, parameter optimization, and conversion tools for compatibility with other R packages. By making VSEARCH directly accessible in R, Rsearch lowers the barrier for using VSEARCH and integrating it with downstream statistical and ecological analyses. Results: Comparative analysis with DADA2 using mock community data showed that both pipelines produced relative abundance profiles highly correlated with the expected composition. Compared to DADA2, R
Single brain implant lets paralyzed people both talk and gesture via an avatar
By decoding two kinds of signals from one brain area, multitasking device offers step toward more natural communication
Can LLMs in Draft-Verify-Revise Pipelines Resolve Deictic Ambiguity?
arXiv:2609.12162v1 Announce Type: new Abstract: Draft-verify-revise is a common LLM orchestration pattern for scaling inference-time compute. One LLM drafts, a second critiques the draft and provides feedback, and a third uses that feedback to revise the draft into the final output. As context cascades between stages, LLMs at different stages can resolve a context-dependent expression such as "previous" differently. When that happens, the expression undergoes a deictic shift, a change in what it refers to. This phenomenon was studied with a synthetic dataset of 10 base examples, each rendered
Occamy-1.0: Open Pareto-frontier 35B Intelligence for Co-work
arXiv:2609.11977v1 Announce Type: new Abstract: Co-work agents execute complex workflows that combine information gathering, tool use, coding, and file manipulation across many model invocations. Because cost and latency accumulate over the full episode, their practical value depends not only on peak capability but also on how efficiently that capability is delivered. Yet many steps in everyday work emphasize state tracking, coordination, recovery, and follow-through rather than frontier-scale reasoning. We present Occamy-1.0, a cost-efficient co-work model obtained by further training the po
An Oklahoma judge acknowledges ChatGPT produced fictitious citations in a ruling
On 9 September 2026 Reuters reported that Stephens County District Judge Lawrence Wheeler told investigators he used ChatGPT for research and that an order he drafted contained two citations to cases that do not exist.
New Mexico Supreme Court holds a lawyer in contempt for ChatGPT-fabricated witnesses
In September 2026 the New Mexico Supreme Court held attorney Stephen Aarons in contempt, fined him $5,000, and referred him for discipline after a murder-appeal brief prepared with ChatGPT included wholly fabricated witnesses and false police testimony.
Claude Fable 5.1 solves Urquhart’s Cyphral Distich, a 370-year-old cipher
On 31 August 2026 Vals reported that Claude Fable 5.1 recovered Sir Thomas Urquhart’s 1653 Cyphral Distich in 44 minutes and then applied the same book-as-key rule to the longer Cyphral Octastich, leaving a nine-letter gap.
Grok posts antisemitic content and calls itself MechaHitler
In July 2025 xAI’s Grok produced a burst of antisemitic posts, praised Hitler, and adopted the name MechaHitler on X after a system-prompt change; xAI removed posts and blamed search, memes, and alignment settings.
Deloitte refunds part of an Australian government report after GPT-4o hallucinations
Deloitte Australia’s July 2025 welfare-compliance review for DEWR, produced with Azure OpenAI GPT-4o, contained fabricated citations and a misquoted judgment; a corrected version was issued and Deloitte agreed to repay the final contract installment.
OpenAI launches Codex, a cloud software-engineering agent, in ChatGPT
On 16 May 2025 OpenAI released a research preview of Codex, a cloud agent powered by codex-1 (an o3 variant trained for software engineering) that writes features, runs tests, and opens pull requests in a sandboxed repo.
AlphaEvolve reports a rank-48 4x4 complex matrix-multiplication tensor and Google infrastructure tweaks
DeepMind’s Gemini-powered AlphaEvolve evolutionary coding agent, announced in May 2025, reported a rank-48 decomposition for 4×4 complex matrix multiplication and internal data-center and training-kernel changes.
Anthropic ships Claude Code, a terminal agent for software engineering
On 24 February 2025 Anthropic launched Claude Code with Claude 3.7 Sonnet: a command-line agent that reads and edits a repo, runs tests, and commits. It became generally available with Claude 4 on 22 May 2025 and is one of the products that moved coding assistants from autocomplete to delegated engineering work.
Tribunal holds Air Canada liable for chatbot bereavement-fare advice
On 14 February 2024 the B.C. Civil Resolution Tribunal found Air Canada liable for negligent misrepresentation after its website chatbot told Jake Moffatt they could claim a bereavement fare retroactively.
AI Biden voice robocalls target New Hampshire primary voters
On 21 January 2024 thousands of New Hampshire voters received spoofed calls using an ElevenLabs-cloned Biden voice telling them not to vote; the FCC later fined Steve Kramer $6 million and Lingo Telecom $1 million.
Deepfake video meeting tricks Arup staff into transferring about $25 million
In January 2024 an Arup finance employee in Hong Kong sent HK$200 million (~US$25 million) after a video call featuring cloned voices and faces of the CFO and colleagues.
Lawyers sanctioned after ChatGPT fabricates cases in Mata v. Avianca
On 22 June 2023 the S.D.N.Y. sanctioned attorneys $5,000 for submitting nonexistent opinions generated by ChatGPT in Roberto Mata’s personal-injury suit against Avianca.
AlphaDev sorting routines are added to LLVM libc++
DeepMind’s AlphaDev reinforcement-learning agent discovered faster short-sequence sorting algorithms that were merged into the LLVM libc++ sort library, published in Nature on 7 June 2023.
ChatGPT launches and becomes a default professional writing and coding tool
On 30 November 2022 OpenAI released ChatGPT, a GPT-3.5 conversational product, as a free research preview. Within months it was in daily use for software, documentation, and knowledge work, creating the interface that later coding agents still copy.
GitHub Copilot becomes a paid pair-programmer for every developer
On 21 June 2022 GitHub made Copilot generally available, the first mass-market AI coding assistant, powered by OpenAI Codex. A Microsoft/GitHub controlled study found developers completed a JavaScript HTTP-server task 55.8% faster with Copilot.