Digg - AI news, before it trends

Dwarkesh Patel, host of the Dwarkesh Podcast, argues against defining intelligence as the ability to achieve goals across domains since it would rank leaders like Trump, Xi, and Putin highest

Dwarkesh Patel challenges the common AI definition of intelligence as the ability to achieve goals across domains. That framing would incorrectly place political leaders like Trump, Xi, and Putin above scientists and abstract reasoners.

Ken Griffin, Citadel founder, says AI agents now complete complex market analysis and deal structuring tasks in hours or days that previously required teams of master’s and PhD holders working weeks or months

Ken Griffin says AI agents at Citadel now finish complex market analysis and deal structuring in hours or days. Tasks that once took teams of master's- and PhD-level professionals weeks or months are being completed this way.

xAI completed training of its 1.5 trillion parameter Grok V9 foundation model as a major upgrade over the current 0.5 trillion parameter Grok V8 version

xAI has completed training of its 1.5 trillion parameter Grok V9 model, a major upgrade from the current 0.5 trillion parameter V8. The next steps include supplemental training on Cursor data, followed by supervised fine-tuning and reinforcement learning, with public release expected in three to four weeks.

Mustafa Suleyman, Microsoft AI CEO, predicts AI will reach human-level performance on most professional tasks within 18 months and fully automate white-collar roles such as accounting, legal, marketing, and project management

Microsoft AI CEO Mustafa Suleyman predicts AI will reach human-level performance on most professional tasks within 18 months, automating white-collar roles such as accounting, legal, marketing and project management. Researchers including Gary Marcus have challenged the timeline with a $100,000 wager and cited ongoing technical barriers in AI robustness.

Venture capitalist Chamath Palihapitiya warns that consulting firms like PwC and Accenture integrating Anthropic or OpenAI models directly enable those providers to capture usage data and fund competing services

Consulting firms like PwC and Accenture risk handing AI providers an edge when they embed OpenAI or Anthropic models directly into client work. The providers can then analyze usage patterns to develop competing services.

Richard Ngo, an independent AI researcher focused on AGI safety, states Effective Altruism’s largest real-world effect has been to empower Anthropic by supplying narratives, funding, and talent rather than its intended impact routes

Independent researcher Richard Ngo argues that Effective Altruism's largest real-world effect has been channeling funding, narratives, and talent to Anthropic instead of the movement's intended impact paths. The claim centers on how these pipelines have supported the AI lab's growth.

Claude Code flushes its KV cache after inactivity, forcing context reconstruction from tokens alone and reducing effectiveness in a way that matches Memento research results.

Claude Code becomes noticeably less effective after idle periods because it clears its KV cache. The model must then reconstruct prior context from tokens alone rather than continuing from preserved internal state.

Jason Liu, OpenAI developer experience engineer on the Codex team, shares draft blog post detailing techniques and primitives for daily Codex use in knowledge work such as presentations and voice notes

An OpenAI Codex team engineer has shared concrete methods for applying the tool to routine knowledge tasks including presentations and voice notes. The draft guide focuses on daily primitives that leverage recent features like durable threads and document review surfaces.

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

Matthew Yglesias proposes tying regulatory approval for new data centers to a requirement that operators reserve 20% of GPUs for low-cost use. The set-aside would target low-income users and AI agents to expand access to scarce high-performance hardware.

Figure's F.03 humanoid robot sorts 12,732 packages at 2.83 seconds each in a livestreamed warehouse demonstration, nearly matching human worker Aime who handled 12,924 packages at 2.79 seconds per package

The robot kept sorting autonomously through day six for 17,444 additional packages, reaching a cumulative 149,000 over 119 hours of continuous runtime.

Fyodor Urnov, professor of molecular therapeutics at Berkeley MCB, prompted Edison Scientific's deep literature mode to generate a 17-page gene biology review PDF in 25 minutes and called the output flawless

Berkeley professor Fyodor Urnov used Edison Scientific's AI tool to produce a full 17-page gene biology review in 25 minutes. The output covered a gene he knows well and was judged flawless on review.

Grok Imagine Generates Anime Video Of Girl Awakening Star Energy

Logan Kilpatrick, focused on Google AI Studio and Gemini API, posts that the model is the product as researchers debate whether full systems now outweigh isolated models

Google engineer Logan Kilpatrick posted that the underlying model itself is now the core product for AI offerings. Replies from researchers and builders split between continued model advances and the rising role of agent behaviors plus learning loops for differentiation.

Google AI Overview Responds As Viltrumite To Invincible Search

Odyssey ML will present one or two world model releases during a livestream tomorrow at 9:30 AM PT to demonstrate their direction and growing capabilities amid ongoing GPU scaling

Odyssey ML plans to show progress on general world models by releasing one or two new versions in a public livestream. The event starts tomorrow at 9:30 AM PT and will highlight the lab's expanding capabilities alongside continued GPU scaling.

AI researchers discuss efficiency advantages of language models over vision models, noting the latter are roughly 1000 times smaller thanks to language's high-density compression via compositional semantics

Language models compress information far more efficiently than vision models due to compositional semantics. Researchers observed that this makes typical vision models roughly 1000 times smaller, an insight that emerged during analysis of how long chain-of-thought reasoning remains coherent.

Chamath Palihapitiya claims on the All-In Podcast that Taiwan will lose its strategic importance in semiconductor manufacturing within 18 months as the industry reaches 1-2 nanometer nodes

Investors and analysts maintain that TSMC continues to handle most advanced chip production for major firms like NVIDIA and Apple. New domestic U.S. fabs are still years away from matching that scale or challenging Taiwan's position in semiconductors.

Posts on X reject viral claims that a new AI data center in Utah would raise the state's nighttime temperatures by 20 to 28 degrees as implausible given the limited scale of heat release from such facilities

Posts on X have rejected claims that an AI data center in Utah would raise the state's nighttime temperatures by 20 to 28 degrees. Comparisons to large wildfires at 2200 degrees Fahrenheit show no such statewide effects, with data centers releasing far less heat.

Nous Research releases Hermes Agent v0.14.0 on GitHub

Nous Research delivered a large-scale update to Hermes Agent with version 0.14.0 now on GitHub. The release covers 808 commits across 1,393 files, adds multi-agent Kanban boards and search tools, and resolves 545 issues.

Pope Leo XIV approves creation of an Interdicasterial Commission on Artificial Intelligence in a May 12 rescript to examine the technology's accelerating adoption and effects on humanity

Pope Leo XIV has created a Vatican commission to assess artificial intelligence's effects on people and society. The body draws from seven Vatican institutions and follows his recent signing of an AI encyclical scheduled for release in the coming weeks.

A Meta data scientist claims current AI lacks capability to displace existing roles but projects a need to advance several abstraction layers higher within roughly one year, citing uncertainty for new field entrants

A Meta data scientist says current AI tools still cannot replace most existing tech roles yet predicts workers will need to operate several abstraction layers higher within roughly one year. New entrants face added uncertainty as career paths and required skills shift rapidly.

Anthropic retires Claude 3 Opus in January 2026

Anthropic retired Claude 3 Opus in January 2026 but continues to offer access on claude.ai and through a special API. The company has approved every submitted application for that API, unlike other recent models that were fully removed.

Redis creator Salvatore Sanfilippo thanks DeepSeek AI

Redis creator Salvatore Sanfilippo publicly thanked DeepSeek AI for making advanced tools and concepts available to wider audiences. The post draws attention to upcoming V4.1 and V4.1-Vision releases expected to expand model access and capabilities.

Caitlin Kalinowski discusses AI hardware boom on Lenny's Podcast

Hardware expert Caitlin Kalinowski describes how the AI industry is shifting focus to physical systems like robots, which face supply-chain limits and potential memory price increases. She shared these insights in a Lenny's Podcast interview drawing from her roles at Apple, Meta, and OpenAI.

Greg Brockman shares Codex demo on personal messages

OpenAI's Codex analyzed three years of personal text messages and generated summaries that incorporated direct quotes, producing an emotionally resonant result for the user. Mac users can run the same process after granting access permissions to their message history.

Opus 4.7 model in Claude Code v2.1.143 fetches and filters 1350 Pokemon entries in 11 seconds

Claude's Opus 4.7 model executed a bash command to pull 1350 Pokemon entries from the PokeAPI and filter them for names ending in 'aw'. Using a 1M context window, the task completed in 11 seconds and returned the two matching results.

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

Max Novendstern links Thiel optimism to Ethereum and OpenAI

Max Novendstern connected Peter Thiel's idea of indefinite optimism to Vitalik Buterin's addition of a virtual machine on Ethereum and OpenAI's decision to scale transformer models. The thread frames both as major technical bets made without assured outcomes, with later replies extending the point to stochastic hardware designs.

xAI SuperGrok Integrates With Hermes Agent For Seamless Access

Nando de Freitas unveils interventional SFT for agent models

A simple change to supervised fine-tuning reduces how often language-model agents reinforce their own incorrect beliefs on factual tasks. The method trains only on observation tokens while ignoring action tokens and outperformed standard fine-tuning across dozens of prompts.

Dwarkesh Patel, host of the Dwarkesh Podcast, argues against defining intelligence as the ability to achieve goals across domains since it would rank leaders like Trump, Xi, and Putin highest

Dwarkesh Patel challenges the common AI definition of intelligence as the ability to achieve goals across domains. That framing would incorrectly place political leaders like Trump, Xi, and Putin above scientists and abstract reasoners.

Ken Griffin, Citadel founder, says AI agents now complete complex market analysis and deal structuring tasks in hours or days that previously required teams of master’s and PhD holders working weeks or months

Ken Griffin says AI agents at Citadel now finish complex market analysis and deal structuring in hours or days. Tasks that once took teams of master's- and PhD-level professionals weeks or months are being completed this way.

xAI completed training of its 1.5 trillion parameter Grok V9 foundation model as a major upgrade over the current 0.5 trillion parameter Grok V8 version

xAI has completed training of its 1.5 trillion parameter Grok V9 model, a major upgrade from the current 0.5 trillion parameter V8. The next steps include supplemental training on Cursor data, followed by supervised fine-tuning and reinforcement learning, with public release expected in three to four weeks.

Mustafa Suleyman, Microsoft AI CEO, predicts AI will reach human-level performance on most professional tasks within 18 months and fully automate white-collar roles such as accounting, legal, marketing, and project management

Microsoft AI CEO Mustafa Suleyman predicts AI will reach human-level performance on most professional tasks within 18 months, automating white-collar roles such as accounting, legal, marketing and project management. Researchers including Gary Marcus have challenged the timeline with a $100,000 wager and cited ongoing technical barriers in AI robustness.

Venture capitalist Chamath Palihapitiya warns that consulting firms like PwC and Accenture integrating Anthropic or OpenAI models directly enable those providers to capture usage data and fund competing services

Consulting firms like PwC and Accenture risk handing AI providers an edge when they embed OpenAI or Anthropic models directly into client work. The providers can then analyze usage patterns to develop competing services.

Richard Ngo, an independent AI researcher focused on AGI safety, states Effective Altruism’s largest real-world effect has been to empower Anthropic by supplying narratives, funding, and talent rather than its intended impact routes

Independent researcher Richard Ngo argues that Effective Altruism's largest real-world effect has been channeling funding, narratives, and talent to Anthropic instead of the movement's intended impact paths. The claim centers on how these pipelines have supported the AI lab's growth.

Claude Code flushes its KV cache after inactivity, forcing context reconstruction from tokens alone and reducing effectiveness in a way that matches Memento research results.

Claude Code becomes noticeably less effective after idle periods because it clears its KV cache. The model must then reconstruct prior context from tokens alone rather than continuing from preserved internal state.

Jason Liu, OpenAI developer experience engineer on the Codex team, shares draft blog post detailing techniques and primitives for daily Codex use in knowledge work such as presentations and voice notes

An OpenAI Codex team engineer has shared concrete methods for applying the tool to routine knowledge tasks including presentations and voice notes. The draft guide focuses on daily primitives that leverage recent features like durable threads and document review surfaces.

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

Matthew Yglesias proposes tying regulatory approval for new data centers to a requirement that operators reserve 20% of GPUs for low-cost use. The set-aside would target low-income users and AI agents to expand access to scarce high-performance hardware.

Figure's F.03 humanoid robot sorts 12,732 packages at 2.83 seconds each in a livestreamed warehouse demonstration, nearly matching human worker Aime who handled 12,924 packages at 2.79 seconds per package

The robot kept sorting autonomously through day six for 17,444 additional packages, reaching a cumulative 149,000 over 119 hours of continuous runtime.

Fyodor Urnov, professor of molecular therapeutics at Berkeley MCB, prompted Edison Scientific's deep literature mode to generate a 17-page gene biology review PDF in 25 minutes and called the output flawless

Berkeley professor Fyodor Urnov used Edison Scientific's AI tool to produce a full 17-page gene biology review in 25 minutes. The output covered a gene he knows well and was judged flawless on review.

Grok Imagine Generates Anime Video Of Girl Awakening Star Energy

Logan Kilpatrick, focused on Google AI Studio and Gemini API, posts that the model is the product as researchers debate whether full systems now outweigh isolated models

Google engineer Logan Kilpatrick posted that the underlying model itself is now the core product for AI offerings. Replies from researchers and builders split between continued model advances and the rising role of agent behaviors plus learning loops for differentiation.

Google AI Overview Responds As Viltrumite To Invincible Search

Odyssey ML will present one or two world model releases during a livestream tomorrow at 9:30 AM PT to demonstrate their direction and growing capabilities amid ongoing GPU scaling

Odyssey ML plans to show progress on general world models by releasing one or two new versions in a public livestream. The event starts tomorrow at 9:30 AM PT and will highlight the lab's expanding capabilities alongside continued GPU scaling.

AI researchers discuss efficiency advantages of language models over vision models, noting the latter are roughly 1000 times smaller thanks to language's high-density compression via compositional semantics

Language models compress information far more efficiently than vision models due to compositional semantics. Researchers observed that this makes typical vision models roughly 1000 times smaller, an insight that emerged during analysis of how long chain-of-thought reasoning remains coherent.

Chamath Palihapitiya claims on the All-In Podcast that Taiwan will lose its strategic importance in semiconductor manufacturing within 18 months as the industry reaches 1-2 nanometer nodes

Investors and analysts maintain that TSMC continues to handle most advanced chip production for major firms like NVIDIA and Apple. New domestic U.S. fabs are still years away from matching that scale or challenging Taiwan's position in semiconductors.

Posts on X reject viral claims that a new AI data center in Utah would raise the state's nighttime temperatures by 20 to 28 degrees as implausible given the limited scale of heat release from such facilities

Posts on X have rejected claims that an AI data center in Utah would raise the state's nighttime temperatures by 20 to 28 degrees. Comparisons to large wildfires at 2200 degrees Fahrenheit show no such statewide effects, with data centers releasing far less heat.

Nous Research releases Hermes Agent v0.14.0 on GitHub

Nous Research delivered a large-scale update to Hermes Agent with version 0.14.0 now on GitHub. The release covers 808 commits across 1,393 files, adds multi-agent Kanban boards and search tools, and resolves 545 issues.

Pope Leo XIV approves creation of an Interdicasterial Commission on Artificial Intelligence in a May 12 rescript to examine the technology's accelerating adoption and effects on humanity

Pope Leo XIV has created a Vatican commission to assess artificial intelligence's effects on people and society. The body draws from seven Vatican institutions and follows his recent signing of an AI encyclical scheduled for release in the coming weeks.

A Meta data scientist claims current AI lacks capability to displace existing roles but projects a need to advance several abstraction layers higher within roughly one year, citing uncertainty for new field entrants

A Meta data scientist says current AI tools still cannot replace most existing tech roles yet predicts workers will need to operate several abstraction layers higher within roughly one year. New entrants face added uncertainty as career paths and required skills shift rapidly.

Anthropic retires Claude 3 Opus in January 2026

Anthropic retired Claude 3 Opus in January 2026 but continues to offer access on claude.ai and through a special API. The company has approved every submitted application for that API, unlike other recent models that were fully removed.

Redis creator Salvatore Sanfilippo thanks DeepSeek AI

Redis creator Salvatore Sanfilippo publicly thanked DeepSeek AI for making advanced tools and concepts available to wider audiences. The post draws attention to upcoming V4.1 and V4.1-Vision releases expected to expand model access and capabilities.

Caitlin Kalinowski discusses AI hardware boom on Lenny's Podcast

Hardware expert Caitlin Kalinowski describes how the AI industry is shifting focus to physical systems like robots, which face supply-chain limits and potential memory price increases. She shared these insights in a Lenny's Podcast interview drawing from her roles at Apple, Meta, and OpenAI.

Greg Brockman shares Codex demo on personal messages

OpenAI's Codex analyzed three years of personal text messages and generated summaries that incorporated direct quotes, producing an emotionally resonant result for the user. Mac users can run the same process after granting access permissions to their message history.

Opus 4.7 model in Claude Code v2.1.143 fetches and filters 1350 Pokemon entries in 11 seconds

Claude's Opus 4.7 model executed a bash command to pull 1350 Pokemon entries from the PokeAPI and filter them for names ending in 'aw'. Using a 1M context window, the task completed in 11 seconds and returned the two matching results.

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