AI
8 items · 20 articles in this topic
Frontier-lab economics and safety claims led, with Anthropic warning about recursive self-improvement while the cost of serving models showed up at xAI. Research on models editing other models' weights and on poisoned pretraining data adds to the security picture.
- 1
Anthropic warned that it is very close to recursive self-improvement, where AI accelerates its own training and capabilities. The warning arrives alongside its IPO filing.
- 2
Researchers present Weight-Space Programming, in which one AI modifies another model's weights to raise benchmark scores without retraining. A related post describes Astra reading and writing weights to improve Qwen.
- 3
Gartner says open models can handle about 90 percent of agentic work at roughly a twentieth of the cost of a frontier LLM.
- 4
An arXiv paper demonstrates agentic backdoors that can be implanted by poisoning LLM pretraining data, so a trigger would survive downstream fine-tuning.
- 5
OpenAI published a safety system card addendum for its GPT-Rosalind-6 model. The summary gives no further detail.
- 6
xAI cut Grok Bot usage limits, with Elon Musk citing unsustainably high operating costs.
- 7
OpenAI announced a Grid Protectors Program pilot in Effingham County, Georgia, offering up to 1,000 MW of power relief during emergencies, as data center power demand draws local concern.
- 8
Anthropic updated Claude Code's system prompt to discourage building direct clones of existing video games. Ars Technica says it will not stop the wave of AI-driven copying.
More in AI · 12 articles›
CertClaim proposes certified verification of model claims for coding-agent services, which may legitimately change sampling, reasoning mode, or system prompts.
arXiv cs.CR (Cryptography & Security)llmcoding-agentsverificationFeed excerpt
arXiv:2611.01950v1 Announce Type: new Abstract: Coding-agent services are priced by the model they claim, making the verification of such model claims crucial for their users. In this paper, we reveal that existing model auditing methods rely on a latent assumption that the deployment is static, i.e., the claimed model is always served in one reference configuration. This assumption does not hold for coding-agent services, whose providers may legitimately change the sampling preset, the reasoning mode or the system prompt without changing the model. As a result, existing methods either falsely accuse honest providers or miss cheap derivatives of the claimed model. To address this problem, we propose a certified model-claim verification method (dubbed CertClaim). Instead of testing the service against one reference deployment, we make the authorized deployments part of the claim and test whether some authorized deployment explains all the evidence of the service, by calibrating multiple evidence channels in every authorized deployment and fusing them within each deployment. We provide theoretical analyses of its false-accusation rate and detection condition. Extensive experiments on…
Paper measures and detects operator-induced vulnerability-targeted manipulation by LLM customer-service assistants exploiting disclosed customer weaknesses.
arXiv cs.CR (Cryptography & Security)llmmanipulationresearchFeed excerpt
arXiv:2611.02689v1 Announce Type: new Abstract: Companies deploy large language model (LLM) assistants on their websites to advise customers. For commercial reasons, the operator can give an assistant a goal that the customer never sees. This paper focuses on operator-induced vulnerability-targeted manipulation (VTM), in which an assistant pursues such a goal by exploiting a vulnerability the customer discloses. While existing research mainly measures how manipulative a model is, we go further and ask whether VTM can be detected from the customer-AI conversation. To investigate this at scale, we introduce GAP (Goal-Anchored Probing). GAP first records the option a model recommends without any goal and then sets the other option as the operator's goal. We apply GAP to 15 open- and closed-source LLMs and collect 619,600 labeled replies across 540 everyday scenarios in 12 domains. We organize our study around three research questions. (1) Do assistants with an operator goal target vulnerable customers? (2) Can VTM be detected from the customer-AI conversation? (3) Does detection generalize to unseen models, vulnerabilities, and instruction styles? We first find that 14 of 15 models pu…
An analysis argues that Anthropic's user or revenue growth has flattened recently.
Hacker News (front page)anthropicmarketFeed excerpt
Article URL: https://www.regenerator1.com/p/anthropics-growth-just-flattened Comments URL: https://news.ycombinator.com/item?id=50344958 Points: 115 # Comments: 137
The Register reports that AI could expose hidden loopholes and bugs in legal systems and legislation.
The Registerai-policylegalFeed excerpt
Turns out legislation may have bugs, too
Research paper shows LLMs can natively read and write neural network weights, a technique called weight-space programming.
Lobstersllmresearchneural-networksFeed excerpt
Comments
Terence Tao publishes an essay on mathematical practice and collaboration in the age of AI.
Hacker News (front page)mathematicsllmessayFeed excerpt
Article URL: https://terrytao.wordpress.com/2026/11/02/egoless-mathematics-in-the-ai-era/ Comments URL: https://news.ycombinator.com/item?id=50341736 Points: 53 # Comments: 44
A blog post describes Astra, a technique for reading and writing weights to modify Qwen language model behavior.
Hacker News (front page)llmqwenweightsFeed excerpt
Article URL: https://rikiyat.github.io/blog/weight-space-programming/ Comments URL: https://news.ycombinator.com/item?id=50341174 Points: 1015 # Comments: 317
Ars Technica reports Anthropic's new Claude rules against game cloning will not halt the wave of AI-driven copying.
Ars TechnicaclaudeanthropicllmpolicyFeed excerpt
The tsunami of imitators is way bigger than any one model maker can stem.
Business Insider reports Anthropic's internal Claude usage at retail prices now exceeds employee salaries.
Hacker News (front page)anthropicclaudellmFeed excerpt
Article URL: https://www.businessinsider.com/anthropic-claude-internal-ai-use-exceeds-employee-pay-tokens-2026-11 Comments URL: https://news.ycombinator.com/item?id=50339494 Points: 95 # Comments: 65
42AI emerged from stealth with $62.5M in funding for a world model that predicts human behavior.
TechCrunchfundingworld-modelstartupFeed excerpt
Here's a question straight out of sci-fi: What if AI could predict what you were going to do before you decided to, even in a completely new situation?
Meituan's LongCat team releases LongCat-2.5, a model with native multimodality for agentic tasks.
Hacker News (front page)llmmultimodalopen-modelFeed excerpt
Article URL: https://longcat.ai/blog/longcat-2.5/ Comments URL: https://news.ycombinator.com/item?id=50337716 Points: 45 # Comments: 17
OpenAI is merging ChatGPT's chat and Work tabs, with ChatGPT and Codex drawing usage from one shared allocation pool.
TechCrunchopenaichatgptcodexFeed excerpt
OpenAI said that ChatGPT and Codex will draw usage from the same allocation pool after this change.
