OpenAI launches ChatGPT personal finance preview for Pro users

OpenAI has started testing a personal finance tool in ChatGPT for Pro users in the United States. The feature connects to bank and brokerage accounts through Plaid, letting users view spending breakdowns and ask questions about their finances in plain language.

HashiCorp co-founder Mitchell Hashimoto warns of AI psychosis

Tech companies risk creating systems that appear healthy on local metrics while becoming globally fragile as they rush to deploy AI agents for rapid automation. HashiCorp co-founder Mitchell Hashimoto highlighted this pattern, noting it echoes earlier cloud infrastructure shifts that favored quick recovery over resilient design.

Clawpatch releases 0.1.0 for semantic codebase mapping

Clawpatch 0.1.0 maps codebases into semantic feature slices that include entrypoints, owned files, tests, and trust boundaries. The CLI invokes the Codex provider to review each slice and stores findings with severity, confidence, evidence, and fix recommendations. Supported commands cover init for project detection across Node.js, TypeScript, Go, Rust, and SwiftPM, plus map, review, fix, and report functions. The release follows experiments that ran roughly 100 Codex agents to scan commits and PRs.

Owain Evans identifies Negation Neglect in Qwen3.5-397B

Large language models can adopt false claims as true after fine-tuning on documents that explicitly label them false. Researcher Owain Evans documented this pattern, called Negation Neglect, in Qwen3.5-397B and similar models.

AI wealth surge widens widest San Francisco tech gap

Roughly 10,000 employees and founders at AI firms including OpenAI, Anthropic, xAI, Nvidia and Meta have reached retirement-level wealth above 20 million dollars. This has produced the widest compensation gap recorded in San Francisco tech over the past five years, while most other engineers remain below 500,000 dollars annually amid continued layoffs.

OpenAI Codex team fixes GPT-5.5 performance issues

OpenAI's Codex team identified and fixed two issues that reduced GPT-5.5 performance over the preceding 48 hours. The model is now operating stably with usage limits scheduled to reset after overnight monitoring.

Geoffrey Irving departs AISI to found AI alignment nonprofit

Geoffrey Irving is leaving the UK AI Security Institute to launch a new nonprofit focused on AI alignment research. He is returning to the Bay Area after roughly two years leading work on scalable alignment and model transparency.

Scott Bessent announces first US-China AI safety talks

The United States and China will hold the first formal AI safety talks of President Trump’s second term. Treasury Secretary Scott Bessent announced the discussions from Beijing, with the focus on protocols to keep advanced AI models away from hackers and terrorists.

Eric Jang releases AutoGo codebase reconstructing AlphaGo

Eric Jang released AutoGo, an open codebase that rebuilds AlphaGo using current reinforcement learning frameworks. It trains agents from scratch with Monte Carlo Tree Search, neural network evaluation, and self-play under Tromp-Taylor rules.

Y Combinator CEO Garry Tan says AI dismantles socialism

Y Combinator CEO Garry Tan argues that AI undermines socialism by generating abundance and giving people direct access to resources without intermediaries. His post highlights how the technology reduces scarcity, eliminates the need for victim protection, and bypasses political allocation channels.

xAI enables Grok subscribers to connect to Hermes Agent

Grok subscribers can now authenticate directly with Hermes Agent using xAI OAuth. The open-source agent from Nous Research runs persistently on computers or servers and connects to WhatsApp, Discord, Telegram, and Signal.

Anthropic resets Claude rate limits for all users

Anthropic restored full quotas for all Claude users by resetting the five-hour and weekly rate limits. The change took effect Friday after the company acquired GPUs from xAI, allowing continued use over the weekend.

X open-sources For You feed recommendation algorithm

Developers and researchers can now inspect the code that ranks posts in X's For You feed. X released the full source for the system on GitHub in the xai-org repository.

FutureSim benchmark evaluates AI agents on continual learning

FutureSim introduces a benchmark that tests how well frontier AI agents update predictions as new information arrives over time. It feeds models like GPT-5.5 sequential news on events such as the Seattle Seahawks Super Bowl odds and Nepal's prime minister selection, tracking forecast changes and accuracy scores.

Founder Tricks Recruiters With Prompt Injection In LinkedIn Bio

Dany Bittel reconstructs blueberry with 3D Gaussian splatting

A detailed 3D model of a blueberry was produced using Gaussian splatting from 6,120 photographs. The reconstruction contains 1.48 million splats and captures the fruit's textured surface along with interior details.

OpenAI integrates ChatGPT and Codex for shared workflows

OpenAI is linking ChatGPT and Codex so users can combine conversation context with code repositories in shared workflows. Codex now runs as a command center on chatgpt.com, orchestrating parallel agents for tasks such as feature builds, refactors, migrations, and CI/CD.

Dan Shipper demonstrates OpenAI Codex piano teaching app

Dan Shipper built a macOS piano teaching app using OpenAI Codex that converts natural language prompts into playable chord sequences. The interface shows code snippets for chords like Fmaj7 alongside a piano roll and virtual keyboard that highlights and plays notes in real time.

Databricks AI Research forms team for AI agent evaluation

Databricks AI Research is forming a team to measure and improve AI agents that operate on large-scale enterprise data. The group will build systems to turn evaluation results into better performance during development, training, and deployment for complex tasks in biotech and finance.

Hudson River Trading expands LLM team in HRT AI Labs

Hudson River Trading is expanding its LLM team to build models from scratch that support a significant share of its trading activity. The firm is hiring engineers across the full training pipeline from pre-training to post-training.

OpenAI releases Codex preview in ChatGPT mobile apps

OpenAI has added a Codex preview to ChatGPT on iOS and Android. Users can start coding tasks, review outputs, and approve steps from a phone while the agent keeps running on a linked laptop or dev machine.

Center for Shared AI Prosperity launches to press policymakers

A new Center for Shared AI Prosperity has formed to urge U.S. policymakers to address large-scale economic changes expected from advanced AI. It focuses on expanding safety nets, restructuring labor markets, and redesigning taxes to share AI benefits more broadly, drawing on data showing most Americans view the system as rigged.

Marc Benioff guest-hosts All-In Podcast episode

Salesforce CEO Marc Benioff discussed how AI will transform SaaS categories and the implications of the Trump-Xi summit on trade and Taiwan's chips during his All-In Podcast appearance. The episode also touched on OpenAI potentially suing Apple and a new multi-sensory AI model.

Figure AI robots reach third day of continuous autonomous operation

Figure AI humanoid robots have now operated continuously without human input for three straight days in a live demonstration. They perform work tasks around the clock using only neural network control and will run until a hardware failure occurs.

Anthropic Mythos AI uncovers macOS M5 kernel vulnerabilities

Researchers used Anthropic's Mythos AI to identify the first public macOS kernel memory corruption vulnerabilities on M5 silicon. The flaws were chained into a privilege-escalation exploit that bypassed Memory Integrity Enforcement after five days and roughly $35,000 in API costs.

Jane Street reveals Texas AI data center with 4,032 GPUs

Jane Street built a new AI data center in Texas to power its trading technology with thousands of GPUs. It contains 4,032 GPUs in 56 racks, linked by 8,000 km of fiber and using liquid cooling.

xAI completes training run for Grok V9 model

xAI finished an internal training run for its Grok V9 foundation model with 1.5 trillion parameters. Early results look strong before the next phase adds data from the Cursor code editor.

Vivian Balakrishnan presents Raspberry Pi AI assistant

Singapore Cabinet Minister Vivian Balakrishnan demonstrated building a personal AI assistant on a Raspberry Pi at the AI Engineer Singapore conference. He presented as a NanoClaw power user during the opening day of talks streamed live on YouTube.

FrontierCS releases FrontierSmith for open-ended coding data

FrontierCS released FrontierSmith, a system that turns closed-ended coding problems into open-ended tasks through mutation and filtering to create runnable environments. Models trained on this synthetic data beat those trained on human-curated sets across multiple benchmarks for long-horizon coding agents.

Greg Isenberg forecasts AI agents within 12 months

AI agents that learn from screen recordings and memory systems could soon handle full jobs after just one week of observation. Such agents are forecast to emerge within 12 months, with current examples already scanning X discussions to generate daily news feeds that improve through ongoing use.

OpenAI launches ChatGPT personal finance preview for Pro users

OpenAI has started testing a personal finance tool in ChatGPT for Pro users in the United States. The feature connects to bank and brokerage accounts through Plaid, letting users view spending breakdowns and ask questions about their finances in plain language.

HashiCorp co-founder Mitchell Hashimoto warns of AI psychosis

Tech companies risk creating systems that appear healthy on local metrics while becoming globally fragile as they rush to deploy AI agents for rapid automation. HashiCorp co-founder Mitchell Hashimoto highlighted this pattern, noting it echoes earlier cloud infrastructure shifts that favored quick recovery over resilient design.

Clawpatch releases 0.1.0 for semantic codebase mapping

Clawpatch 0.1.0 maps codebases into semantic feature slices that include entrypoints, owned files, tests, and trust boundaries. The CLI invokes the Codex provider to review each slice and stores findings with severity, confidence, evidence, and fix recommendations. Supported commands cover init for project detection across Node.js, TypeScript, Go, Rust, and SwiftPM, plus map, review, fix, and report functions. The release follows experiments that ran roughly 100 Codex agents to scan commits and PRs.

Owain Evans identifies Negation Neglect in Qwen3.5-397B

Large language models can adopt false claims as true after fine-tuning on documents that explicitly label them false. Researcher Owain Evans documented this pattern, called Negation Neglect, in Qwen3.5-397B and similar models.

AI wealth surge widens widest San Francisco tech gap

Roughly 10,000 employees and founders at AI firms including OpenAI, Anthropic, xAI, Nvidia and Meta have reached retirement-level wealth above 20 million dollars. This has produced the widest compensation gap recorded in San Francisco tech over the past five years, while most other engineers remain below 500,000 dollars annually amid continued layoffs.

OpenAI Codex team fixes GPT-5.5 performance issues

OpenAI's Codex team identified and fixed two issues that reduced GPT-5.5 performance over the preceding 48 hours. The model is now operating stably with usage limits scheduled to reset after overnight monitoring.

Geoffrey Irving departs AISI to found AI alignment nonprofit

Geoffrey Irving is leaving the UK AI Security Institute to launch a new nonprofit focused on AI alignment research. He is returning to the Bay Area after roughly two years leading work on scalable alignment and model transparency.

Scott Bessent announces first US-China AI safety talks

The United States and China will hold the first formal AI safety talks of President Trump’s second term. Treasury Secretary Scott Bessent announced the discussions from Beijing, with the focus on protocols to keep advanced AI models away from hackers and terrorists.

Eric Jang releases AutoGo codebase reconstructing AlphaGo

Eric Jang released AutoGo, an open codebase that rebuilds AlphaGo using current reinforcement learning frameworks. It trains agents from scratch with Monte Carlo Tree Search, neural network evaluation, and self-play under Tromp-Taylor rules.

Y Combinator CEO Garry Tan says AI dismantles socialism

Y Combinator CEO Garry Tan argues that AI undermines socialism by generating abundance and giving people direct access to resources without intermediaries. His post highlights how the technology reduces scarcity, eliminates the need for victim protection, and bypasses political allocation channels.

xAI enables Grok subscribers to connect to Hermes Agent

Grok subscribers can now authenticate directly with Hermes Agent using xAI OAuth. The open-source agent from Nous Research runs persistently on computers or servers and connects to WhatsApp, Discord, Telegram, and Signal.

Anthropic resets Claude rate limits for all users

Anthropic restored full quotas for all Claude users by resetting the five-hour and weekly rate limits. The change took effect Friday after the company acquired GPUs from xAI, allowing continued use over the weekend.

X open-sources For You feed recommendation algorithm

Developers and researchers can now inspect the code that ranks posts in X's For You feed. X released the full source for the system on GitHub in the xai-org repository.

FutureSim benchmark evaluates AI agents on continual learning

FutureSim introduces a benchmark that tests how well frontier AI agents update predictions as new information arrives over time. It feeds models like GPT-5.5 sequential news on events such as the Seattle Seahawks Super Bowl odds and Nepal's prime minister selection, tracking forecast changes and accuracy scores.

Founder Tricks Recruiters With Prompt Injection In LinkedIn Bio

Dany Bittel reconstructs blueberry with 3D Gaussian splatting

A detailed 3D model of a blueberry was produced using Gaussian splatting from 6,120 photographs. The reconstruction contains 1.48 million splats and captures the fruit's textured surface along with interior details.

OpenAI integrates ChatGPT and Codex for shared workflows

OpenAI is linking ChatGPT and Codex so users can combine conversation context with code repositories in shared workflows. Codex now runs as a command center on chatgpt.com, orchestrating parallel agents for tasks such as feature builds, refactors, migrations, and CI/CD.

Dan Shipper demonstrates OpenAI Codex piano teaching app

Dan Shipper built a macOS piano teaching app using OpenAI Codex that converts natural language prompts into playable chord sequences. The interface shows code snippets for chords like Fmaj7 alongside a piano roll and virtual keyboard that highlights and plays notes in real time.

Databricks AI Research forms team for AI agent evaluation

Databricks AI Research is forming a team to measure and improve AI agents that operate on large-scale enterprise data. The group will build systems to turn evaluation results into better performance during development, training, and deployment for complex tasks in biotech and finance.

Hudson River Trading expands LLM team in HRT AI Labs

Hudson River Trading is expanding its LLM team to build models from scratch that support a significant share of its trading activity. The firm is hiring engineers across the full training pipeline from pre-training to post-training.

OpenAI releases Codex preview in ChatGPT mobile apps

OpenAI has added a Codex preview to ChatGPT on iOS and Android. Users can start coding tasks, review outputs, and approve steps from a phone while the agent keeps running on a linked laptop or dev machine.

Center for Shared AI Prosperity launches to press policymakers

A new Center for Shared AI Prosperity has formed to urge U.S. policymakers to address large-scale economic changes expected from advanced AI. It focuses on expanding safety nets, restructuring labor markets, and redesigning taxes to share AI benefits more broadly, drawing on data showing most Americans view the system as rigged.

Marc Benioff guest-hosts All-In Podcast episode

Salesforce CEO Marc Benioff discussed how AI will transform SaaS categories and the implications of the Trump-Xi summit on trade and Taiwan's chips during his All-In Podcast appearance. The episode also touched on OpenAI potentially suing Apple and a new multi-sensory AI model.

Figure AI robots reach third day of continuous autonomous operation

Figure AI humanoid robots have now operated continuously without human input for three straight days in a live demonstration. They perform work tasks around the clock using only neural network control and will run until a hardware failure occurs.

Anthropic Mythos AI uncovers macOS M5 kernel vulnerabilities

Researchers used Anthropic's Mythos AI to identify the first public macOS kernel memory corruption vulnerabilities on M5 silicon. The flaws were chained into a privilege-escalation exploit that bypassed Memory Integrity Enforcement after five days and roughly $35,000 in API costs.

Jane Street reveals Texas AI data center with 4,032 GPUs

Jane Street built a new AI data center in Texas to power its trading technology with thousands of GPUs. It contains 4,032 GPUs in 56 racks, linked by 8,000 km of fiber and using liquid cooling.