Digg - AI news, before it trends

Dwarkesh Patel distinguishes intelligence as the ability to achieve goals across wide domains from raw power and ranks leaders like Donald Trump, Xi Jinping, and Vladimir Putin at the top

Discussions on advanced AI separate intelligence as abstract reasoning from the capacity to wield power and achieve goals. Political leaders such as Donald Trump and Vladimir Putin illustrate high real-world influence that does not require strong performance on scientific or conceptual tasks, while current AI training targets economically rewarded skills.

Ken Griffin of Citadel says AI agents complete complex market analysis and deal structuring tasks in hours or days that previously took teams of master's- and PhD-level finance professionals weeks or months

Citadel CEO Ken Griffin reports that AI agents now complete complex market analysis and deal structuring in hours or days. These tasks previously required teams of master's- and PhD-level professionals working for weeks or months.

xAI completed training of its 1.5 trillion parameter Grok V9 foundation model, which Elon Musk described as a major upgrade over the 0.5 trillion parameter Grok V8 now running as version 4.3

xAI has completed training of its 1.5 trillion parameter Grok V9 model, a major upgrade from the current 0.5 trillion parameter V8 version. The team will next add Cursor data through supplemental training and fine-tuning before a public release in three to four weeks.

Mustafa Suleyman, Microsoft AI CEO, predicts AI will reach human-level performance on most professional tasks within 18 months resulting in full automation of white-collar roles such as accounting and legal work

Microsoft AI CEO Mustafa Suleyman predicts AI will reach human-level performance on most professional tasks within 18 months. This timeline points to full automation of white-collar roles such as accounting, legal work, marketing, and project management.

Chamath Palihapitiya states that consulting firms like PwC and Accenture risk displacement by feeding usage data to Anthropic and OpenAI while those providers launch competing services

Consulting firms like PwC and Accenture risk displacement as OpenAI and Anthropic use client data from direct integrations to launch competing services. Chamath Palihapitiya and others point to routing requests via intermediaries like Databricks to retain data control.

Richard Ngo, an independent AI researcher and philosopher focused on AGI safety, alignment, and governance, stated that Effective Altruism’s largest real-world effect has been to empower Anthropic through narratives, funding, and talent pipelines.

AI researcher Richard Ngo says Effective Altruism's main real-world effect has been building up Anthropic instead of its original goals. Support flowed through funding, narratives, and talent pipelines to the AI lab.

Dimitris Papailiopoulos reports that Claude Code flushes its KV cache after periods of inactivity, aligning with Memento paper findings on reconstruction challenges upon resumption.

Claude Code automatically flushes its KV cache after periods of inactivity, causing noticeably reduced effectiveness upon resumption. This forces the model to reconstruct prior context solely from tokens, consistent with findings from the Memento paper.

OpenAI's Jason Liu shares a draft GitHub post outlining methods to maximize Codex use by shifting from coding to broader knowledge work

OpenAI developer Jason Liu has published methods for applying the Codex AI model to a wider range of daily tasks beyond coding. His draft GitHub post describes core primitives and workstreams that users can adopt for tasks like note-taking and document creation.

Matthew Yglesias proposes conditioning approval of new data centers on operators setting aside 20% of GPUs as affordable compute for low-income agents

Matthew Yglesias proposed requiring new data centers to reserve 20% of GPUs for low-income compute as a condition for approval. Responses have compared the approach to existing rules for affordable housing in new developments.

Figure's F.03 humanoid robot sorts packages at 2.83 seconds each against warehouse worker Aime's 2.79-second pace while accumulating 149,000 packages over 119 hours of autonomous operation

A Figure humanoid robot ran autonomously for 119 hours and sorted 149,000 packages in a warehouse test. It handled each package in 2.83 seconds, narrowly behind a human worker's 2.79-second average during the same shift.

Fyodor Urnov, Professor of Molecular Therapeutics at Berkeley, tested Edison Scientific's deep literature mode by requesting a gene biology review and received a flawless 17-page PDF within 25 minutes

AI tools are generating detailed scientific literature reviews in minutes rather than through slower traditional journal processes. Berkeley professor Fyodor Urnov received a flawless 17-page gene biology PDF from Edison Scientific's deep literature mode in 25 minutes.

Grok Imagine Generates Anime Video Of Girl Awakening Star Energy

Google's Logan Kilpatrick posts that the model is the product as replies examine models operating as components inside agent-driven systems with continual learning and user experiences

Google engineer Logan Kilpatrick posted that AI models are the core product. Replies argue this view is outdated, with value now shifting to models that operate invisibly inside agent systems with continual learning and refined user experiences.

Google AI Overview Responds As Viltrumite To Invincible Search

Odyssey ML will livestream one or two world model releases tomorrow at 9:30 AM PT to clarify their trajectory and capabilities as GPU scaling continues

Odyssey ML will release one or two new world models during a livestream tomorrow. The 9:30 AM PT event will show how the models are progressing as the company scales up its GPU resources.

Gavin Leech, UK-based AI researcher and co-founder of the consultancy Arb, notes vision models are roughly 1000 times smaller than text models owing to language's data compression through compositional semantics and abstractions

AI vision models stay roughly 1000 times smaller than text models because language compresses complex information through abstractions and compositional semantics. Replies in the thread proposed image-based chain-of-thought reasoning as one route to more efficient visual processing.

Chamath Palihapitiya claims on the All-In podcast that Taiwan will lose strategic importance in semiconductor manufacturing within 18 months due to U.S. advances in 1-2 nanometer process nodes

US progress on domestic 1-2 nanometer semiconductor production is not expected to erode Taiwan's strategic role in chip manufacturing within the next 18 months. Industry observers note that TSMC still controls the vast majority of advanced process nodes, packaging, and supply chains used by firms like NVIDIA and Apple.

Social media users reject claims that a new Utah AI data center would raise nighttime temperatures by 20 to 28 degrees as physically implausible

Social media users are rejecting viral claims that a new Utah AI data center would raise nighttime temperatures by 20 to 28 degrees statewide. They note that even intense wildfires reaching 2200°F do not produce comparable regional warming effects.

NousResearch releases Hermes Agent v0.14.0 on May 16, 2026, with 808 commits, 633 merged pull requests, changes to 1,393 files, and 545 closed issues

NousResearch released Hermes Agent v0.14.0 as a foundation update with extensive code revisions. The version adds multi-agent Kanban boards, goal integration, X and Grok search, and a Codex CLI tool.

Pope Leo XIV approves an Interdicasterial Commission on Artificial Intelligence via May 12 rescript to examine AI advances and effects on humanity

Pope Leo XIV established a new Vatican interdicasterial commission to examine artificial intelligence advances and their effects on people and society. The move follows his recent signing of an AI encyclical set for release in the coming weeks.

Approximately 10,000 employees and founders at OpenAI, Anthropic, xAI, and Nvidia have each accumulated over $20 million in wealth as AI compresses traditional tech employment disruption cycles for software engineers

Roughly 10,000 employees and founders at OpenAI, Anthropic, xAI, and Nvidia have each accumulated over $20 million in wealth. Many other software engineers now plan to move to higher abstraction layers in their roles within the next year.

Anthropic retired its March 2024 Claude 3 Opus model in January 2026 while keeping access available on claude.ai and through an API form that approves all requests, unlike other decommissioned models

After retiring Claude 3 Opus in January 2026, Anthropic continues to offer the March 2024 model on claude.ai. Users can also request API access through a form that has approved all submissions so far, unlike other decommissioned models.

Salvatore Sanfilippo, creator of Redis, thanked DeepSeek AI directly for making artificial intelligence and its underlying concepts available to a broader audience

Salvatore Sanfilippo, creator of the Redis database, thanked DeepSeek AI for making artificial intelligence tools and concepts available to a wider audience. His post highlighted the company's upcoming V4.1 and V4.1-Vision releases.

Lenny Rachitsky releases Lenny’s Podcast episode featuring Caitlin Kalinowski on her Apple, Meta, and OpenAI hardware leadership and AI moving into physical domains like humanoid robots

Caitlin Kalinowski’s experience leading hardware at Apple, Meta, and OpenAI illustrates how AI development is moving from software into physical systems such as humanoid robots. In the new Lenny’s Podcast episode she discusses her career building consumer electronics and robotics teams.

Yuchen Jin, who builds AI systems and products at Databricks, reports worsening GPU shortages as H100 prices exceed three-year-old levels and on-demand availability disappears due to major labs' multi-year contracts

University researchers and independent developers now struggle to access current-generation GPUs for their work. H100 prices exceed levels from three years ago and on-demand capacity has disappeared after major labs signed multi-year supply contracts.

Riley Brown documents Codex on macOS generating an emotionally moving summary from three years of personal text messages after instructions to incorporate direct quotes

Codex produced an emotionally moving summary after processing three years of personal text messages and including direct quotes from the conversations. The approach runs on macOS once the user grants file-access permissions.

Claude Code v2.1.143 running Opus 4.7 with a 1M context window executes a Bash command to fetch data for 1350 Pokémon from the PokeAPI and filters names ending in 'aw' within 11 seconds

Claude Code with a 1M context window accurately performs API data retrieval and filtering that direct prompts often get wrong. In a demo it fetched records for all 1,350 Pokémon from PokeAPI and returned names ending in 'aw' within 11 seconds.

LangSmith Engine Enables Always-On Self-Improvement For AI Agents

Max Novendstern tweets link indefinite optimism to Vitalik Buterin's Ethereum virtual machine and OpenAI's scaled transformers while contrasting Peter Thiel's strategic assessment

Max Novendstern's tweet links indefinite optimism to Vitalik Buterin's choice to build a virtual machine into Ethereum and OpenAI's decision to scale transformers. The post contrasts this view with Peter Thiel's assessment of the mindset as a weak strategy and extends it to stochastic hardware.

xAI SuperGrok Integrates With Hermes Agent For Seamless Access

Dwarkesh Patel distinguishes intelligence as the ability to achieve goals across wide domains from raw power and ranks leaders like Donald Trump, Xi Jinping, and Vladimir Putin at the top

Discussions on advanced AI separate intelligence as abstract reasoning from the capacity to wield power and achieve goals. Political leaders such as Donald Trump and Vladimir Putin illustrate high real-world influence that does not require strong performance on scientific or conceptual tasks, while current AI training targets economically rewarded skills.

Ken Griffin of Citadel says AI agents complete complex market analysis and deal structuring tasks in hours or days that previously took teams of master's- and PhD-level finance professionals weeks or months

Citadel CEO Ken Griffin reports that AI agents now complete complex market analysis and deal structuring in hours or days. These tasks previously required teams of master's- and PhD-level professionals working for weeks or months.

xAI completed training of its 1.5 trillion parameter Grok V9 foundation model, which Elon Musk described as a major upgrade over the 0.5 trillion parameter Grok V8 now running as version 4.3

xAI has completed training of its 1.5 trillion parameter Grok V9 model, a major upgrade from the current 0.5 trillion parameter V8 version. The team will next add Cursor data through supplemental training and fine-tuning before a public release in three to four weeks.

Mustafa Suleyman, Microsoft AI CEO, predicts AI will reach human-level performance on most professional tasks within 18 months resulting in full automation of white-collar roles such as accounting and legal work

Microsoft AI CEO Mustafa Suleyman predicts AI will reach human-level performance on most professional tasks within 18 months. This timeline points to full automation of white-collar roles such as accounting, legal work, marketing, and project management.

Chamath Palihapitiya states that consulting firms like PwC and Accenture risk displacement by feeding usage data to Anthropic and OpenAI while those providers launch competing services

Consulting firms like PwC and Accenture risk displacement as OpenAI and Anthropic use client data from direct integrations to launch competing services. Chamath Palihapitiya and others point to routing requests via intermediaries like Databricks to retain data control.

Richard Ngo, an independent AI researcher and philosopher focused on AGI safety, alignment, and governance, stated that Effective Altruism’s largest real-world effect has been to empower Anthropic through narratives, funding, and talent pipelines.

AI researcher Richard Ngo says Effective Altruism's main real-world effect has been building up Anthropic instead of its original goals. Support flowed through funding, narratives, and talent pipelines to the AI lab.

Dimitris Papailiopoulos reports that Claude Code flushes its KV cache after periods of inactivity, aligning with Memento paper findings on reconstruction challenges upon resumption.

Claude Code automatically flushes its KV cache after periods of inactivity, causing noticeably reduced effectiveness upon resumption. This forces the model to reconstruct prior context solely from tokens, consistent with findings from the Memento paper.

OpenAI's Jason Liu shares a draft GitHub post outlining methods to maximize Codex use by shifting from coding to broader knowledge work

OpenAI developer Jason Liu has published methods for applying the Codex AI model to a wider range of daily tasks beyond coding. His draft GitHub post describes core primitives and workstreams that users can adopt for tasks like note-taking and document creation.

Matthew Yglesias proposes conditioning approval of new data centers on operators setting aside 20% of GPUs as affordable compute for low-income agents

Matthew Yglesias proposed requiring new data centers to reserve 20% of GPUs for low-income compute as a condition for approval. Responses have compared the approach to existing rules for affordable housing in new developments.

Figure's F.03 humanoid robot sorts packages at 2.83 seconds each against warehouse worker Aime's 2.79-second pace while accumulating 149,000 packages over 119 hours of autonomous operation

A Figure humanoid robot ran autonomously for 119 hours and sorted 149,000 packages in a warehouse test. It handled each package in 2.83 seconds, narrowly behind a human worker's 2.79-second average during the same shift.

Fyodor Urnov, Professor of Molecular Therapeutics at Berkeley, tested Edison Scientific's deep literature mode by requesting a gene biology review and received a flawless 17-page PDF within 25 minutes

AI tools are generating detailed scientific literature reviews in minutes rather than through slower traditional journal processes. Berkeley professor Fyodor Urnov received a flawless 17-page gene biology PDF from Edison Scientific's deep literature mode in 25 minutes.

Grok Imagine Generates Anime Video Of Girl Awakening Star Energy

Google's Logan Kilpatrick posts that the model is the product as replies examine models operating as components inside agent-driven systems with continual learning and user experiences

Google engineer Logan Kilpatrick posted that AI models are the core product. Replies argue this view is outdated, with value now shifting to models that operate invisibly inside agent systems with continual learning and refined user experiences.

Google AI Overview Responds As Viltrumite To Invincible Search

Odyssey ML will livestream one or two world model releases tomorrow at 9:30 AM PT to clarify their trajectory and capabilities as GPU scaling continues

Odyssey ML will release one or two new world models during a livestream tomorrow. The 9:30 AM PT event will show how the models are progressing as the company scales up its GPU resources.

Gavin Leech, UK-based AI researcher and co-founder of the consultancy Arb, notes vision models are roughly 1000 times smaller than text models owing to language's data compression through compositional semantics and abstractions

AI vision models stay roughly 1000 times smaller than text models because language compresses complex information through abstractions and compositional semantics. Replies in the thread proposed image-based chain-of-thought reasoning as one route to more efficient visual processing.

Chamath Palihapitiya claims on the All-In podcast that Taiwan will lose strategic importance in semiconductor manufacturing within 18 months due to U.S. advances in 1-2 nanometer process nodes

US progress on domestic 1-2 nanometer semiconductor production is not expected to erode Taiwan's strategic role in chip manufacturing within the next 18 months. Industry observers note that TSMC still controls the vast majority of advanced process nodes, packaging, and supply chains used by firms like NVIDIA and Apple.

Social media users reject claims that a new Utah AI data center would raise nighttime temperatures by 20 to 28 degrees as physically implausible

Social media users are rejecting viral claims that a new Utah AI data center would raise nighttime temperatures by 20 to 28 degrees statewide. They note that even intense wildfires reaching 2200°F do not produce comparable regional warming effects.