Tech News

It’s a new day at Digg. Fresh stories are clustering now. 85% complete.

Video Signals

#1360

Seedance 2.5 AI Video Model Releases with Stunning Demo

6h ago69K868198

#593

OpenAI Tests Bidirectional Voice Model Bidi 1 for ChatGPT

5h ago72K1K279

#184

Midjourney Stands Out For Unique Atmospheric AI Images And Animations

8h ago76K1K316

#171

Alibaba Launches HappyHorse 1.1 Video Synthesis Model via API

5h ago32K16243

#571

Liberty Mutual's Vlad Barbalat Discusses AI Impact on Equity Multiples

1h ago30K80120

#684

GLM-5.2 Runs Fast on Modal With Parallel Code Agents

7h ago5.8K7113

#684

QodoAI Ships Cross-Repo Review To Detect Bugs Across Multiple Repos

48m ago3.9K189

#1861

Roemmele Urges Weaponizing AI To Fix Technical Debt Y2K-Style

9h ago15K5412

#1499

Gen2Balance Uses Generative Images to Balance Long-Tailed Video Action Recognition

1h ago918253

#1238

Gemini App Adds Soccer Templates For Custom Trading Cards And Plushies

38m ago2.1K234

Google DeepMind's Séb Krier argues polarized AI narratives fill an existential or religious void for advocatesHenry Shevlin countered that grand AI narratives track historical reality

2h5.2k11114

#667

Top post: @AndrewCurran_ “@sebkrier It's different for everyone. For me it's because I had forsaken the world. I regret it deeply, but I did. I could see no way to change our path, and finally, I gave up. Then ten years later, using GPT for the first time in 2022, I realized there was hope.”

Cursor AI announces a partnership with SpaceX to train a new model at its Compile keynoteNo release date or technical specifications have been disclosed

4h37.2k83037

RISING LIKES830

#530

Top post: @beffjezos “@ericzakariasson It's out of this world?”

New github stars (48hrs)

Orchestrates multiple AI coding agents across terminal, browser, and phone with shared live sessions

omnigent-ai/omnigent4.5k10h

Implements a domain-agnostic framework for autonomous AI agents with ReAct loops, multi-LLM routing, reasoning graphs and MCP integration.

441

684

Maksym Andriushchenko of ELLIS Institute says GLM-5.2 hits a 74.8% hallucination rate on HalluHard without web searchGPT-5.4-Thinking registered a 46.8% rate under the same conditions.

4h2.5k423

FASTEST CLIMBING+5

#1207

Top post: @scaling01 “hallucination final boss”

7NEW

Founder Ryan Petersen argues remote work with young kids is impractical, while ml.school's Santiago Valdarrama says offices destroy productivityValdarrama argued that office spaces systematically diminish employee motivation.

3h14.3k11918

#395

Top post: @typesfast “@HarryStebbings 🤔”

AI developer Teortaxes and research engineer Alexander Doria dispute claims that compute advantages dictate AI capability milestonesThey point to China's GLM agent deployment as evidence.

4h8.1k926

#501

Top post: @Dorialexander “real question always was: what are you modeling exactly? compute is just the way to get there.”

Meta's Lucas Beyer and Cody Blakeney argue modern plotting tools lack the precision of ggplot and SeabornModern tools quickly deliver 80% but struggle with polish.

4h3.4k476

#72

Top post: @giffmana “@code_star Yep. But also, vibe-plotting gets you to 80% in 20% of the time, which is great for exploring results and presentation styles quickly. But, so far, I've had quite a lot of trouble getting the last 20% to a really good plot.”

106

A California nonprofit is attempting to block construction of a semiconductor fabrication plant in New York StateThe opposition is led by Los Angeles-based equity coordinators.

3h976340

#492

Top post: @bscholl “@beffjezos They’re just anti human”

111

Rohan Anil, CoreAutoAI co-founder and former Gemini pretraining lead, proposes standardizing the AI training stack starting with tokenizersSupporters say standardization improves training efficiency as data sources diverge

4h5.2k621

#102

Top post: @Dorialexander “@_arohan_ Training efficiency (even more significant now that labs are not at all using the same data sources).”

120

Interconnects editor Florian Brand argues fully automated AI content lacks value, making minimal manual editing the new differentiatorDeveloper @xlr8harder uses AI for structuring before manual rewriting

4h3.3k444

#1778

Top post: @xeophon “My assumption is that everyone I talk to uses AI as much (or more) than I do, so they can also generate the same stuff. When stuff is fully generated, why would I read it or listen? I can do it myself. Even refining AI slop is putting you at the top 1% these days fwiw”

130

Simo Ryu, Stable Diffusion LoRA creator, argues tech firms gain a competitive advantage by avoiding certain researchersSusan Zhang countered that firms manage talent acquisition independently.

4h66360

#84

Top post: @suchenzang “@cloneofsimo not for you and i to worry about, kitten corporations will figure it out without needing to collude against individuals”

14NEW

Probook raises $40 million from a16z and Sequoia for its AI-powered home service dispatch softwareIt unifies technician dispatch with customer management workflows.

23m10100

#1089

Top post: @georgeprobook “Announcing $40M in funding from @sequoia and @a16z. I grew up pressure washing houses in upstate New York. Six summers in the truck, missing calls from the top of a ladder because I couldn't hear my phone ring. My co-founder Ben grew up plumbing with his dad in San Francisco. We entered home services because we grew up in it. I spent a summer inside TR Miller, one of the best-run HVAC, plumbing and electrical shops in the country, and their workflows were still raw. The first wave of AI in the trades was point solutions. One tool for voice, one for outbound, one for follow-up. Operators ended up with five vendors and more overhead. Probook is the AI operating system that helps you add points to your EBITDA and run your entire customer experience as one connected system, built around dispatch. Many of America’s largest home service brands run on Probook today. Summers Plumbing booked 2,542 jobs their first month with us, zero human intervention. Anthony PHCE ran 20% more revenue per job on a 50% leaner team. Del-Air more than doubled productivity from 10 to 22 techs per dispatcher. To everyone who trusted us early: thank you. Back to work. We're hiring.”

158

An AI PhD's LeetCode hurdles before joining OpenAI spark debate over hiring tests for machine learning rolesThe hire had thousands of citations and major model contributions.

1h53040

#741

Top comment: @natolambert “@code_star I indeed got a silly ml coding interview in my “job search”. Everyone goes through it.”

162

AI developer Elie Bakouch teases 32B-parameter Playmobil MoE model with no technical details or repository linkThe unverified post drew immediate skepticism from ML researchers.

3h92040

#353

Top post: @nathanbenaich “@eliebakouch 😂”

17NEW

Raj Jayaram proves that multi-vector embeddings are exponentially more expressive than single vectors for similarity searchSingle-vector models require exponentially more dimensions to match performance.

13m16400

#200

Top post: @mrdrozdov “This is a good paper.”

Recent Stars

Github Stars

(7 days)

Today's Highlights

OpenAI partners with Trail of Bits to automatically patch open-source vulnerabilities, merging 37 fixes in its first week

OpenAI tests "Bidi 1," a bidirectional ChatGPT voice model that can speak and listen simultaneously

Seedance 2.5 announced to generate 30-second 4K videos from up to 50 reference inputs

AMP PBC's Anjney Midha claims tech executives over 50 are irrelevant to AI pricing, drawing pushback from Nikesh Arora

Top Stories

Seedance 2.5 announced to generate 30-second 4K videos from up to 50 reference inputsCreator Chubby clarified the tool is not yet released.

OpenAI tests "Bidi 1," a bidirectional ChatGPT voice model that can speak and listen simultaneouslyEarly demonstrations show fluid counting without conversational pauses.

AMP PBC's Anjney Midha claims tech executives over 50 are irrelevant to AI pricing, drawing pushback from Nikesh AroraArora argued curiosity and adaptability matter more than age.

Video Signals

Google DeepMind's Séb Krier argues polarized AI narratives fill an existential or religious void for advocatesHenry Shevlin countered that grand AI narratives track historical reality

Cursor AI announces a partnership with SpaceX to train a new model at its Compile keynoteNo release date or technical specifications have been disclosed

Maksym Andriushchenko of ELLIS Institute says GLM-5.2 hits a 74.8% hallucination rate on HalluHard without web searchGPT-5.4-Thinking registered a 46.8% rate under the same conditions.

Founder Ryan Petersen argues remote work with young kids is impractical, while ml.school's Santiago Valdarrama says offices destroy productivityValdarrama argued that office spaces systematically diminish employee motivation.

AI developer Teortaxes and research engineer Alexander Doria dispute claims that compute advantages dictate AI capability milestonesThey point to China's GLM agent deployment as evidence.

Meta's Lucas Beyer and Cody Blakeney argue modern plotting tools lack the precision of ggplot and SeabornModern tools quickly deliver 80% but struggle with polish.

A California nonprofit is attempting to block construction of a semiconductor fabrication plant in New York StateThe opposition is led by Los Angeles-based equity coordinators.

Rohan Anil, CoreAutoAI co-founder and former Gemini pretraining lead, proposes standardizing the AI training stack starting with tokenizersSupporters say standardization improves training efficiency as data sources diverge

Interconnects editor Florian Brand argues fully automated AI content lacks value, making minimal manual editing the new differentiatorDeveloper @xlr8harder uses AI for structuring before manual rewriting

Simo Ryu, Stable Diffusion LoRA creator, argues tech firms gain a competitive advantage by avoiding certain researchersSusan Zhang countered that firms manage talent acquisition independently.

Probook raises $40 million from a16z and Sequoia for its AI-powered home service dispatch softwareIt unifies technician dispatch with customer management workflows.

An AI PhD's LeetCode hurdles before joining OpenAI spark debate over hiring tests for machine learning rolesThe hire had thousands of citations and major model contributions.

AI developer Elie Bakouch teases 32B-parameter Playmobil MoE model with no technical details or repository linkThe unverified post drew immediate skepticism from ML researchers.

Raj Jayaram proves that multi-vector embeddings are exponentially more expressive than single vectors for similarity searchSingle-vector models require exponentially more dimensions to match performance.

Recent Stars

Github Stars

Today's Highlights

OpenAI partners with Trail of Bits to automatically patch open-source vulnerabilities, merging 37 fixes in its first week

OpenAI tests "Bidi 1," a bidirectional ChatGPT voice model that can speak and listen simultaneously

Seedance 2.5 announced to generate 30-second 4K videos from up to 50 reference inputs

AMP PBC's Anjney Midha claims tech executives over 50 are irrelevant to AI pricing, drawing pushback from Nikesh Arora

Top Stories

Seedance 2.5 announced to generate 30-second 4K videos from up to 50 reference inputsCreator Chubby clarified the tool is not yet released.

OpenAI tests "Bidi 1," a bidirectional ChatGPT voice model that can speak and listen simultaneouslyEarly demonstrations show fluid counting without conversational pauses.

AMP PBC's Anjney Midha claims tech executives over 50 are irrelevant to AI pricing, drawing pushback from Nikesh AroraArora argued curiosity and adaptability matter more than age.

Video Signals

Google DeepMind's Séb Krier argues polarized AI narratives fill an existential or religious void for advocatesHenry Shevlin countered that grand AI narratives track historical reality

Cursor AI announces a partnership with SpaceX to train a new model at its Compile keynoteNo release date or technical specifications have been disclosed

Maksym Andriushchenko of ELLIS Institute says GLM-5.2 hits a 74.8% hallucination rate on HalluHard without web searchGPT-5.4-Thinking registered a 46.8% rate under the same conditions.

Founder Ryan Petersen argues remote work with young kids is impractical, while ml.school's Santiago Valdarrama says offices destroy productivityValdarrama argued that office spaces systematically diminish employee motivation.

AI developer Teortaxes and research engineer Alexander Doria dispute claims that compute advantages dictate AI capability milestonesThey point to China's GLM agent deployment as evidence.

Meta's Lucas Beyer and Cody Blakeney argue modern plotting tools lack the precision of ggplot and SeabornModern tools quickly deliver 80% but struggle with polish.

A California nonprofit is attempting to block construction of a semiconductor fabrication plant in New York StateThe opposition is led by Los Angeles-based equity coordinators.

Rohan Anil, CoreAutoAI co-founder and former Gemini pretraining lead, proposes standardizing the AI training stack starting with tokenizersSupporters say standardization improves training efficiency as data sources diverge

Interconnects editor Florian Brand argues fully automated AI content lacks value, making minimal manual editing the new differentiatorDeveloper @xlr8harder uses AI for structuring before manual rewriting

Simo Ryu, Stable Diffusion LoRA creator, argues tech firms gain a competitive advantage by avoiding certain researchersSusan Zhang countered that firms manage talent acquisition independently.

Probook raises $40 million from a16z and Sequoia for its AI-powered home service dispatch softwareIt unifies technician dispatch with customer management workflows.

An AI PhD's LeetCode hurdles before joining OpenAI spark debate over hiring tests for machine learning rolesThe hire had thousands of citations and major model contributions.

AI developer Elie Bakouch teases 32B-parameter Playmobil MoE model with no technical details or repository linkThe unverified post drew immediate skepticism from ML researchers.

Raj Jayaram proves that multi-vector embeddings are exponentially more expressive than single vectors for similarity searchSingle-vector models require exponentially more dimensions to match performance.

Recent Stars

Github Stars

Yesterday's Top Stories, Jun 22, 2026.

OpenAI partners with Trail of Bits to automatically patch open-source vulnerabilities, merging 37 fixes in its first weekThe initiative uses GPT-5.5-Cyber to secure Python, Go, and cURL.

Sakana AI launches Fugu orchestration platform, but critics dispute its frontier-level performance claimsFugu Ultra coordinates existing frontier models using test-time compute.

SpaceX signs $6.3 billion deal to lease Nvidia GB300 GPUs from Colossus 2 to Reflection AIReflection AI will pay $150 million monthly through 2029

Ai2 launches TMax, an open-source terminal agent whose 27B model matches systems 40 times its size on Terminal Bench 2.0The suite includes a 14,600-environment dataset and complete training weights

Meta invests $900 million in Indian fintech CRED as founder Kunal Shah steps down to lead WhatsApp globallyThe deal values the fintech platform at $4.5 billion.

Finbarr Timbers of the Allen Institute for AI advises candidates to secure other job offers before applying to top-choice labsThe advice follows an engineer's failed nine-round Anthropic interview.

University of Washington PhD student Alisa Liu, joining OpenAI, details her 57-interview job search at frontier labsHer guide notes one referral equaled 40 cold applications.

OpenClaw steward Peter Steinberger reports the non-profit AI agent project just hit its strongest week of performanceIts non-profit model avoids external venture capital agendas

Google DeepMind partners with film studio A24 to develop creative AI tools based on filmmaker feedbackScott Belsky's Labs team is also collaborating.

Dean W. Ball argues the US AI industry has frozen public releases until the Claude Fable 5 impasse is resolvedPolymarket odds for the model's return dropped to 39%39.

Stripe's Patrick Collison launches Stripe Directory, an experimental tool for AI agents to discover payment-supported businessesThe tool provides machine-readable business profiles via a CLI.

US official clarifies that Senator Warner misunderstood a briefing claiming the AI model Mythos breached classified NSA networksThe briefing actually described an internal network red-teaming exercise

Google's Philipp Schmid announces general availability of Gemini Interactions API with a remote Linux sandboxA background parameter supports asynchronous, long-running agent interactions.

20VC founder Harry Stebbings marks his 30th birthday by sharing his decade record of raising $1 billionHis portfolio includes 18 startups that reached unicorn status.

Ghostty creator Mitchell Hashimoto pledges $400,000 to Zig Software Foundation despite its strict anti-AI policyThe donation brings his family's total contribution to $700,000

Micron partners with Anthropic to design AI-optimized memory architectures and invests in its Series H roundMicron will deploy Claude models across its enterprise operations.

Zai_org's open-weights GLM-5.2 ranks third overall on the GDPval-AA v2 benchmark, beating GPT-5.5 on agentic tasksIt scored 1524 Elo, trailing only two Claude models

Academic Emma Pierson argues AI development should be slowed, claiming societal risks outweigh breakthroughs like curing cancerThe essay triggered sharp criticism from venture capitalist Marc Andreessen.

Open-source AI builder @xlr8harder argues that doctors' poor statistical literacy blocks adoption of new diagnostic imaging toolsCreator @eigenrobot agreed, warning of higher public comprehension barriers.

Entropix creator xjdr announces Noumena has integrated GLM 5.2, splitting its API to isolate 1-million context spamThe update makes GLM the default model for ncode.

Ethan Mollick and Peter Steinberger find Sakana AI's Fugu Ultra struggles on complex coding tasks, lagging behind GPT-5.5One complex prompt exhausted Fugu’s five-hour usage quota

Machine learning journal TMLR will implement submission and authorship quotas using quadratic voting to curb a surge in papersSubmissions recently spiked to 500 papers in 19 days.

Peter Wildeford argues Anthropic has trained a Mythos successor projected to top four hours on the METR benchmarkThe New Carlisle data center expanded to 687,000 H100-equivalent GPUs.

Paul Graham compares AI interfaces like Claude Code to early web browsers acting as a new software layerAusten Allred runs most daily tasks via Claude Code.

Coinbase launches pre-IPO perpetual futures contracts for OpenAI and Anthropic for non-US customersThe synthetic contracts are executed entirely on-chain.

Google invests $75 million in studio A24 to build AI film production and distribution toolsThe joint research partnership targets production and distribution workflows.