Digg - AI news, before it trends

Highlights

Andrej Karpathy joins Anthropic to conduct research and development on frontier large language models after prior roles at OpenAI and Tesla

It’s a new day at Digg. Fresh stories are clustering now. 45% complete.

GitHub Stars (7-days)

…ilsonnn/image-blaster3d ago
Converts an input image into 3D meshes, Gaussian splats, and audio files using Claude skills and external AI APIs.
1673
475
878
263
3.1k stars
…tonomous-speedrunning5d ago
Archives experiments of AI agents autonomously tuning optimizers, schedules, and hyperparameters to reach target validation loss in fewest steps on a small LM benchmark.
214
707
1262
65 stars
manaflow-ai/cmux17h ago
Integrates Ghostty into a macOS terminal with vertical tabs, sidebar metadata, OSC notifications, and in-app browser for AI coding agents.
203
339
17.5k stars
sapientinc/HRM-Text20h ago
Pretrains 1B HRM text models with hierarchical reasoning, task completion, and low-compute PrefixLM training.
1047
707
389 stars
earendil-works/pi1d ago
Provides AI agent toolkit with coding CLI, unified LLM API, TUI/web UI libs and Slack bot support.
219
84
51.6k stars

Recent GitHub Stars

…lama/antigravity-demo2h ago
Sets up and runs a CLI demo of an Antigravity agent using a GitHub repo with LlamaParse for unstructured document parsing.
677
3 stars
CapSoftware/Cap4h ago
Records screens in instant or studio modes for quick capture, local editing, and link-based sharing with self-hosting.
1829
19k stars
…-Mos/salesbench-prime5h ago
Simulates and scores LLM sales agents in a deterministic insurance environment with CRM tools, calls, quotes, and 100 fixed leads.
707
3 stars
…blic/nano-world-model6h ago
Trains video world models via diffusion-forcing with a Hydra-configured pipeline for training, evaluation and planning.
1533
435 stars
…eshgoplani/agent-deck6h ago
Manages terminal sessions for AI coding agents in one TUI with forking, orchestration, monitoring, and Docker sandboxing.
1533
2.5k stars

Digg - AI news, before it trends

Highlights

Andrej Karpathy joins Anthropic to conduct research and development on frontier large language models after prior roles at OpenAI and Tesla

Top Stories.

HRM-Text paper from Sapient Intelligence and MIT presents hierarchical recurrent pretraining that lets a 1B model trained on 40B tokens reach 60.7% MMLU, 84.5% GSM8K and other strong scores in one day of compute — A 0.6B TRM variant beats standard 3B-scale Transformers on downstream tasks.

DeepSeek forms a new Harness team in Beijing to build Code Harness from the ground up, an internal code product that may be released publicly as DeepSeek Code — Job openings posted for Harness Product Manager and R&D Engineer.

Google's Gemini 3.5 Flash is based on the Gemini 3 Flash architecture according to its system card, explaining a price increase driven by higher token speeds and interactivity rather than added scale. — It lags leading models by 2-3 months on the main intelligence index.

Will Depue recounts nightmare in which former President Obama joined Anthropic and argues the AI company should not be allowed to have him — Miles Brundage replied to the post with approval.

Shawn 'swyx' Wang says deep research features in AI have lost relevance since models such as o3, as iterative questioning outperforms single-prompt long reports for active learning — Alexander Doria favors conversational informal drafts over formal reports.

S. Somasegar dies at age 59 after leading Microsoft’s Developer Division for 27 years and backing cloud and AI startups at Madrona Venture Group — Microsoft and Madrona confirmed the death of the Seattle investor.

Kyutai Labs and the ELLIS Institute launch Kesai Labs, a Franco-German non-profit open-science organization developing the first fully open self-driving technology stack based in Tübingen and Paris — The lab commits to releasing open weights, data, and code.

Richard Ngo and Toby Ord examine how AI safety efforts produce unintended outcomes including accelerated progress and institutional advantages from safety rhetoric — Ord is senior researcher at Oxford University's AI Governance Initiative.

Theo Browne says earlier Flash versions were strong but calls Flash 3 a decline and Flash 3.5 unsatisfactory in an X thread on AI models and lab principles — Replies also came from Andreas Kirsch and Susan Zhang.

GitHub Stars (7-days)

Recent GitHub Stars

Digg - AI news, before it trends

Highlights

Andrej Karpathy joins Anthropic to conduct research and development on frontier large language models after prior roles at OpenAI and Tesla

Top Stories.

HRM-Text paper from Sapient Intelligence and MIT presents hierarchical recurrent pretraining that lets a 1B model trained on 40B tokens reach 60.7% MMLU, 84.5% GSM8K and other strong scores in one day of compute — A 0.6B TRM variant beats standard 3B-scale Transformers on downstream tasks.

DeepSeek forms a new Harness team in Beijing to build Code Harness from the ground up, an internal code product that may be released publicly as DeepSeek Code — Job openings posted for Harness Product Manager and R&D Engineer.

Google's Gemini 3.5 Flash is based on the Gemini 3 Flash architecture according to its system card, explaining a price increase driven by higher token speeds and interactivity rather than added scale. — It lags leading models by 2-3 months on the main intelligence index.

Will Depue recounts nightmare in which former President Obama joined Anthropic and argues the AI company should not be allowed to have him — Miles Brundage replied to the post with approval.

Shawn 'swyx' Wang says deep research features in AI have lost relevance since models such as o3, as iterative questioning outperforms single-prompt long reports for active learning — Alexander Doria favors conversational informal drafts over formal reports.

S. Somasegar dies at age 59 after leading Microsoft’s Developer Division for 27 years and backing cloud and AI startups at Madrona Venture Group — Microsoft and Madrona confirmed the death of the Seattle investor.

Kyutai Labs and the ELLIS Institute launch Kesai Labs, a Franco-German non-profit open-science organization developing the first fully open self-driving technology stack based in Tübingen and Paris — The lab commits to releasing open weights, data, and code.

Richard Ngo and Toby Ord examine how AI safety efforts produce unintended outcomes including accelerated progress and institutional advantages from safety rhetoric — Ord is senior researcher at Oxford University's AI Governance Initiative.

Theo Browne says earlier Flash versions were strong but calls Flash 3 a decline and Flash 3.5 unsatisfactory in an X thread on AI models and lab principles — Replies also came from Andreas Kirsch and Susan Zhang.

GitHub Stars (7-days)

Recent GitHub Stars

Yesterday's Top Stories, May 19, 2026.

Andrej Karpathy joins Anthropic to conduct research and development on frontier large language models after prior roles at OpenAI and Tesla — He remains committed to education and will resume projects later.

Google announces Gemini 3.5 model series and Gemini Omni through DeepMind at I/O conference, adding video generation and editing from image and audio references. — Sundar Pichai also updated SynthID with new partners including NVIDIA and OpenAI.

Google releases Gemini 3.5 Flash for complex agentic workflows, coding, and reasoning at 289 tokens per second and higher scores than Gemini 3.1 Pro on Terminal-Bench 2.1, GDPval-AA, and MCP Atlas — Large token quotas appeared in Google Cloud Console before wider rollout.

Analysis projects equity stakes and founder pledges at OpenAI and Anthropic will yield $370 billion in philanthropic assets supporting up to $100 billion in annual grants — OpenAI Foundation's stake already surpasses the Gates Foundation by assets.

Google DeepMind expands Project Genie with Street View integration to generate interactive 3D environments anchored to real-world Google Maps locations — Tool is available for testing by Gemini Ultra users.

METR publishes its first Frontier Risk Report assessing whether AI labs could lose control of advanced autonomous agents after testing internal models from Anthropic, Google, Meta, and OpenAI — Ajeya Cotra led the project after joining METR on January 12.

Google’s Gemini 3.5 Flash scores 55 on the Artificial Analysis Intelligence Index, a nine-point gain over Gemini 3 Flash, while leading the intelligence-speed frontier and exceeding 280 output tokens per second — It posts 76.2 percent on Terminal-bench 2.1.

Google DeepMind releases Omni, a video generation model that advances character consistency and editing by turning short clips into reusable face-and-voice avatars — Demonstrations show consistent characters at multiple Disneyland locations with branded items.

Claude announces the public beta of self-hosted sandboxes and research preview of MCP tunnels within Claude Managed Agents, enabling runs inside user infrastructure under existing security controls — Features were presented at Code with Claude London with Cloudflare and Modal support.

Cursor AI releases Composer 2.5, its most powerful model to date, with gains in intelligence, long-task performance, and complex instruction adherence while doubling usage for all users — Cursor plans a larger model trained from scratch with SpaceXAI on 10x compute.

Edison Scientific announces partnership with Incyte to deploy Kosmos AI agent across full drug development pipeline from molecular design through FDA approval — Incyte becomes first company to integrate the system end-to-end.

OpenAI introduces Guaranteed Capacity to let eligible customers secure long-term compute access across supported cloud providers via 1-3 year commitments and spend-linked discounts — Targets organizations running critical AI workloads and production infrastructure.

Antigravity unifies its agentic surfaces under a single platform incorporating Antigravity 2.0 desktop app, CLI, SDK, and IDE for consistent access across environments — Demo builds Street Guesser game using multi-agent Gemini interfaces.

Sam Altman offers $2 million in OpenAI tokens to every startup in the current Y Combinator batch in exchange for equity — Announcement made at live Y Combinator event using retail token values

Andrej Karpathy joins Anthropic pretraining team and will lead a new group that uses Claude to conduct research, according to journalist Alex Heath — One post featured a satirical portrait of him as Pope Dario I.

xAI enables Grok and X Premium subscribers to connect models to OpenClaw starting today for agent chat, image and video generation, and X post search — OpenClaw is an open-source local-first agent with persistent memory.

Gemini Omni Flash generates a 10-second video from a London Eye prompt that accurately depicts the capsule interior, River Thames, Big Ben, and gentle wheel sway — Output closely matches actual footage from the same location.

Andrej Karpathy joins Anthropic as platform posts respond with surprised Pikachu and awakening gorilla memes — Reactions include a shocked Pikachu close-up and jungle gorilla meme.

Thinking Machines Lab announces grant program awarding multiple $100,000 awards plus $25,000 in Tinker credits to researchers advancing human-AI interactivity with proposals due June 19 2026 — Grants support real-time multimodal proactive and safe human-AI collaboration research.

Google launches its biggest Search redesign in 25 years featuring an AI-powered box with Gemini 3.5 Flash for multimodal queries and agentic capabilities — Announced May 19, the update demos micro apps like weekend planners.

Google showcases SynthID watermarking system at I/O with logos of OpenAI, Nvidia, Kakao and ElevenLabs signaling partner adoption — Sundar Pichai also presented Gemini 3.5 Flash benchmarks at the session.

Viktor raises $75 million Series A led by Accel, the largest ever for a Polish-founded company, as its AI coworker operates inside Slack and connects to over 3,000 tools — It reached $15 million annualized recurring revenue in ten weeks.

Granta magazine asked Claude AI whether passages in a Commonwealth Short Story Prize entry were AI-generated and received opposite conclusions on rephrased prompts — Claude said it is not a detector and such checks are weak.

Cerebras places Kimi K2.6, a trillion-parameter model, into enterprise trials running at roughly 1,000 output tokens per second, the highest speed Artificial Analysis has recorded for any frontier model — Benchmarks show 981 tokens per second on 10,000 input tokens.

Gemini-3.5-Flash regains the top spot on the Toolathlon leaderboard after five months with a 56.5 percent Pass@1 score on 108 agent tasks — Gemini variants also hit 67.42 percent on Terminal-Bench 2.0 physics tasks.

Google announces at its I/O conference that monthly tokens processed across its AI surfaces have reached 3.2 quadrillion — Gemini reached over 900 million monthly active users by April 2026.

xAI publishes setup guide for xurl skill connecting Hermes Agents to the X platform for natural language commands — Nous Research, maker of Hermes models, shared the guide.

Parallel launches Index, a platform that lets content owners monitor AI agents' use of their material and receive compensation based on impressions, citations, uniqueness, and task value — Initial partners include The Atlantic and Fortune with over 100,000 daily users.

Steven Adler and Page Hedley launch Guidelight AI Standards to promote safe frontier AI development practices with first standards on control and transparency requirements — Adler and Hedley also co-authored a report on xAI safety for SpaceX investors.

Yesterday's Top Stories, May 19, 2026.

Andrej Karpathy joins Anthropic to conduct research and development on frontier large language models after prior roles at OpenAI and Tesla — He remains committed to education and will resume projects later.

Google announces Gemini 3.5 model series and Gemini Omni through DeepMind at I/O conference, adding video generation and editing from image and audio references. — Sundar Pichai also updated SynthID with new partners including NVIDIA and OpenAI.

Google releases Gemini 3.5 Flash for complex agentic workflows, coding, and reasoning at 289 tokens per second and higher scores than Gemini 3.1 Pro on Terminal-Bench 2.1, GDPval-AA, and MCP Atlas — Large token quotas appeared in Google Cloud Console before wider rollout.

Analysis projects equity stakes and founder pledges at OpenAI and Anthropic will yield $370 billion in philanthropic assets supporting up to $100 billion in annual grants — OpenAI Foundation's stake already surpasses the Gates Foundation by assets.

Google DeepMind expands Project Genie with Street View integration to generate interactive 3D environments anchored to real-world Google Maps locations — Tool is available for testing by Gemini Ultra users.

METR publishes its first Frontier Risk Report assessing whether AI labs could lose control of advanced autonomous agents after testing internal models from Anthropic, Google, Meta, and OpenAI — Ajeya Cotra led the project after joining METR on January 12.

Google’s Gemini 3.5 Flash scores 55 on the Artificial Analysis Intelligence Index, a nine-point gain over Gemini 3 Flash, while leading the intelligence-speed frontier and exceeding 280 output tokens per second — It posts 76.2 percent on Terminal-bench 2.1.

Google DeepMind releases Omni, a video generation model that advances character consistency and editing by turning short clips into reusable face-and-voice avatars — Demonstrations show consistent characters at multiple Disneyland locations with branded items.

Claude announces the public beta of self-hosted sandboxes and research preview of MCP tunnels within Claude Managed Agents, enabling runs inside user infrastructure under existing security controls — Features were presented at Code with Claude London with Cloudflare and Modal support.

Cursor AI releases Composer 2.5, its most powerful model to date, with gains in intelligence, long-task performance, and complex instruction adherence while doubling usage for all users — Cursor plans a larger model trained from scratch with SpaceXAI on 10x compute.

Edison Scientific announces partnership with Incyte to deploy Kosmos AI agent across full drug development pipeline from molecular design through FDA approval — Incyte becomes first company to integrate the system end-to-end.

OpenAI introduces Guaranteed Capacity to let eligible customers secure long-term compute access across supported cloud providers via 1-3 year commitments and spend-linked discounts — Targets organizations running critical AI workloads and production infrastructure.

Antigravity unifies its agentic surfaces under a single platform incorporating Antigravity 2.0 desktop app, CLI, SDK, and IDE for consistent access across environments — Demo builds Street Guesser game using multi-agent Gemini interfaces.

Sam Altman offers $2 million in OpenAI tokens to every startup in the current Y Combinator batch in exchange for equity — Announcement made at live Y Combinator event using retail token values

Andrej Karpathy joins Anthropic pretraining team and will lead a new group that uses Claude to conduct research, according to journalist Alex Heath — One post featured a satirical portrait of him as Pope Dario I.

xAI enables Grok and X Premium subscribers to connect models to OpenClaw starting today for agent chat, image and video generation, and X post search — OpenClaw is an open-source local-first agent with persistent memory.

Gemini Omni Flash generates a 10-second video from a London Eye prompt that accurately depicts the capsule interior, River Thames, Big Ben, and gentle wheel sway — Output closely matches actual footage from the same location.

Andrej Karpathy joins Anthropic as platform posts respond with surprised Pikachu and awakening gorilla memes — Reactions include a shocked Pikachu close-up and jungle gorilla meme.

Thinking Machines Lab announces grant program awarding multiple $100,000 awards plus $25,000 in Tinker credits to researchers advancing human-AI interactivity with proposals due June 19 2026 — Grants support real-time multimodal proactive and safe human-AI collaboration research.