Digg - AI news, before it trends

It’s a new day at Digg. Fresh stories are clustering now. 25% complete.

Github Stars

(7 days)

…ian-spaced-repetition5H AGO AGO
Schedules spaced-repetition reviews for flashcards and notes inside Obsidian.
681
484
2.4k stars
…-group/InferenceBench1D AGO AGO
Evaluates autonomous AI agents optimizing LLM serving like vLLM in open-ended scenarios with quality and integrity gates.
1224
1066
13 stars
manaflow-ai/cmux2D AGO AGO
Integrates Ghostty into a macOS terminal with vertical tabs, sidebar metadata, OSC notifications, and in-app browser for AI coding agents.
203
339
17.6k stars
sapientinc/HRM-Text2D AGO AGO
Pretrains 1B HRM text models with hierarchical reasoning, task completion, and low-compute PrefixLM training.
1047
707
565 stars
earendil-works/pi3D AGO AGO
Provides AI agent toolkit with coding CLI, unified LLM API, TUI/web UI libs and Slack bot support.
219
84
52.4k stars

Digg - AI news, before it trends

Today's Highlights

OpenAI reports an internal reasoning model identified new super-linear constructions for Erdős's planar unit distance problem from 1946

Top Stories

KVCache.ai launches open-source web-based KV Cache Size Calculator for models including DeepSeek V4 Flash, Qwen3, GLM, Kimi, and MiniMax — DeepSeek V4 Flash at 1M tokens needs 2.893 GiB total cache.

François Fleuret, Meta FAIR research scientist and University of Geneva professor, argues that pure AI systems outperform any human-AI collaboration on sophisticated reasoning tasks — Yaroslav Bulatov replied citing comparative advantage under finite AI energy limits.

Gergely Neu, an ML researcher at ICREA and Universitat Pompeu Fabra, introduces Value-Driven Transport, a framework integrating optimal control and reinforcement learning with optimal transport for generative modeling — Thread shows diagrams and update rule evolving samples to target distribution

DeepSeek recruits R&D engineers, product managers, and researchers for its Harness team focused on frontier research, with Beijing positions open to full-time hires and interns — AI commentators on X shared the recruitment post.

Opus 4.7 displays more guarded responses in the official app during extended interactions due to a built-in long conversation reminder — One session paused to reference prior exchanges and weigh strategies.

Github Stars

Digg - AI news, before it trends

Today's Highlights

OpenAI reports an internal reasoning model identified new super-linear constructions for Erdős's planar unit distance problem from 1946

Top Stories

KVCache.ai launches open-source web-based KV Cache Size Calculator for models including DeepSeek V4 Flash, Qwen3, GLM, Kimi, and MiniMax — DeepSeek V4 Flash at 1M tokens needs 2.893 GiB total cache.

François Fleuret, Meta FAIR research scientist and University of Geneva professor, argues that pure AI systems outperform any human-AI collaboration on sophisticated reasoning tasks — Yaroslav Bulatov replied citing comparative advantage under finite AI energy limits.

Gergely Neu, an ML researcher at ICREA and Universitat Pompeu Fabra, introduces Value-Driven Transport, a framework integrating optimal control and reinforcement learning with optimal transport for generative modeling — Thread shows diagrams and update rule evolving samples to target distribution

DeepSeek recruits R&D engineers, product managers, and researchers for its Harness team focused on frontier research, with Beijing positions open to full-time hires and interns — AI commentators on X shared the recruitment post.

Opus 4.7 displays more guarded responses in the official app during extended interactions due to a built-in long conversation reminder — One session paused to reference prior exchanges and weigh strategies.

Github Stars

Yesterday's Top Stories, May 21, 2026.

OpenAI reports an internal reasoning model identified new super-linear constructions for Erdős's planar unit distance problem from 1946 — The model used algebraic number theory methods without math-specific training.

OpenAI solved the unit distance problem in mathematics, prompting discussion on whether the work used informal chain-of-thought reasoning or neurosymbolic tools such as Lean — Attention now turns to AI progress in social science domains.

Modal raises $355 million in Series C funding at a $4.65 billion valuation led by General Catalyst and Redpoint after crossing $200 million in annual recurring revenue — Its platform powers AI training, inference, and sandbox workloads for Anthropic, Meta, and others.

Microsoft ends internal use of Anthropic's Claude coding tool after token-based billing costs prove unsustainable despite its resources — Uber CTO reports exhausting full 2026 AI budget in four months.

Alibaba's Qwen team releases Qwen3.7-Max, its closed-weights flagship model for agent workflows that completes long-horizon tasks like 35-hour kernel optimizations with over 1,000 tool calls — Scored 56.6 on Artificial Analysis Intelligence Index after 4.8-point gain.

Cursor releases Composer 2.5 coding agent scoring 62 on the Artificial Analysis index while posting 10 to 60 times lower costs than Claude Opus 4.7 and GPT-5.5 — Internal test routed all company chats to the model for two days.

Sam Altman, OpenAI co-founder and CEO, asks what problem people most hope AI will solve in the future — Replies suggest curing diseases and automating research.

Turbopuffer reached a $100 million annual run rate in March, 19 months after hitting the $1 million mark, while remaining profitable on less than $1 million in total funding — Its object storage database first powered Cursor and now serves Anthropic and Notion.

Stanford researchers release paper 'A Bitter Lesson for Data Filtering' showing large models improve without data filtering in high-compute data-scarce regimes — Work targets abundant compute and limited data pretraining scenarios.

Every publishes 'After Automation' report detailing Codex and Claude Code automation of coding, writing, design and customer service as staff grows from 4 to 30 — Every co-founder Dan Shipper notes AI raised demand for human work.

ClickUp announces 22% workforce reduction to restructure around a 100x AI-driven organization with AI-centric workflows — CEO DJ Curfew says the business has never been stronger.

Gemini 3.5 Flash records a leading 47.1% score on the APEX-Agents AA leaderboard and outperforms GPT-5.5 by nearly 10 points on Artificial Analysis benchmarks — It runs 4x faster than Gemini 3.1 Pro but costs twice as much.

Cursor, the AI code editor from Anysphere, doubles usage limits for the next month for new users invited onto its Teams plan — CEO Michael Truell announced the offer on X alongside new feature references.

GoodfireAI shares research showing sparse autoencoders tile and shatter curved neural manifolds in large models rather than linear directions, recasting unsupervised discovery as an inverse Ising problem — Visualizations map features across a manifold spanning 1800 to 1998.

ChatGPT adds a direct integration with Microsoft PowerPoint through a new app that lets users generate slides, query deck content, and apply edits without leaving the application — Built by developer Ryan Brewer at chatgpt.com/apps/powerpoint.

OpenAI enabled a locked use feature for Codex that permits the system to control Mac applications from a connected phone even when the device is locked and the screen is off — It requires an Apple authorization plug-in for macOS sessions.

Hark raises over $700 million in Series A funding at a $6 billion post-money valuation to scale its AI foundation models and custom hardware — Round led by Parkway Venture Capital with NVIDIA and AMD participating

Codex releases Appshots for its Mac app, letting users press Command-Command to capture screenshots and text from open windows including off-screen content and attach them to the active thread — Codex Thursday no. 6 update also adds goal command and remote use while locked.

Antigravity tripled the rate limits for Gemini models across all paid tiers and reset weekly Gemini quotas for users — Update covers every subscription level including 3.5 Flash.

White House postpones signing ceremony for executive order on AI and cybersecurity risks from advanced models after review of provisions that could slow US companies — Axios cited internal administration disagreements over the order.

Pangram Labs analysis flags AI-generated text in three 2026 Commonwealth Short Story Prize regional winners plus the 2025 overall winner and a Granta-published story — Literary foundation and Granta maintain the entries are human-written.

Aleksander Madry, a machine learning researcher at OpenAI and MIT professor on leave, announced his departure after nearly three years heading preparedness and safety efforts — He was reassigned last summer to a role focused on AI reasoning.

Greg Brockman, OpenAI president and co-founder, says the model alone is no longer the product in a post addressing standalone AI offerings — Victor Taelin replied that the claim is wrong and signaled he is shorting.

OpenAI will provide $2 million in tokens to every startup in the current Y Combinator batch in exchange for equity — Sam Altman presented the offer at a Y Combinator event.

SpaceX is hiring world-class engineers and physicists for its SpaceXAI team and states that no prior artificial intelligence experience is required — Elon Musk directed applications to ai_eng@spacex.com with three bullet points.

LeRobot releases LeRobot Humanoid, an open robot-learning platform centered on a bipedal robot assembled for approximately $2,500 using 3D-printed parts and off-the-shelf components — Platform supplies hardware designs, software, and simulation tools for full robot development cycle.

CODA reparameterizes memory-bound operations such as norms, RoPE, residuals, and SwiGLU to execute as part of GEMM epilogues in transformer blocks — CuTeDSL implementation reduces kernel launch overhead on GPUs.

Figure AI humanoid robots complete 191 hours of continuous autonomous warehouse operations, processing over 238,000 packages with the Helix-02 system and a pixels-to-actions neural network. — A timelapse showed 30,160 packages handled in 24 equivalent hours.

Grok Build version 0.1.214 rolls out with X search integration and significantly faster web search along with fixes for Windows PowerShell rendering and keybinding errors — Full rollout is planned by end of day.

AI Security Institute published a report on May 21, 2026, assessing oversight mechanisms for advanced AI systems against rapid capability gains and identifying degradation pathways. — Report section records disputed advantages of discrete token reasoning.