Tech News

Today's Highlights

Ramp and Revelio Labs study of 21,559 firms finds high-intensity AI adopters grew headcount by 10% over two years

+6 more

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

Top Stories

Posts:—Clusters:—Next Crawl:—

Anthropic pulls Claude Fable 5 offline one day after redeployment following a successful jailbreak by security researcher PlinyThe model was redeployed following discussions with the US government.

2h4.3k311

#22

Google DeepMind's Pranav Shyam argues short-horizon PPO reduces to bandit problems, but open-source developer Grad disputes the latency impactGrad says 20,000-token reasoning workloads bypass critic bottlenecks.

1h13562

#1525

Top comment: @Grad62304977 “Like my point is for long horizon I would trust the value models less and less and with compute scaling I would prefer a bigger group size or other ideas to improve credit assignment here If we assume the value model is pretty good then surely it should’ve always been what we should’ve been doing and save lots of compute still”

Prime Intellect's kalomaze argues GANs remain ubiquitous because they power the autoencoders used in modern diffusion modelsThe debate counters claims that the technology is obsolete.

1h23530

#598

Top post: @gabriel1 “@willdepue GANs are hella satisfying”

Watchlist

Video Signals

#1857

AI Subagent Scripts 11 Hand Transformations Using Omni and Lyria 3

2h ago850104

#1311

Robot Juggling Skills Progress From Three To Five Balls In Two Attempts

1h ago1.1K00

#1260

Perplexity CEO: Power Users Now Consume Millions in AI Compute

2h ago97300

#1893

SuperNori Debuts Proactive AI Agent For Family Caretakers

3h ago2.2K00

Wireless EMG Bionic Hand Delivers Independent Finger Control

41m ago25300

Elie Bakouch says Claude's safety filters blocked a routine request to retrieve an Artifact link from a past sessionThe filter flagged the cross-session request as a policy violation

2h10020

#1138

Top post: @eliebakouch “@xeophon you know the worst part, the color was like blue for claude and orange for codex and just asked in another convo to switch it and send the artifact link 😭”

Google DeepMind's Pranav Shyam defines short context horizons at 32k tokens or fewer and long context at 64k or moreLonger horizons require complex multi-round tool calls and code execution.

1h1820

#1525

Top comment: @Grad62304977 “@recurseparadox but whats the intuition here that tool calls and environments have a big effect here. As in its a clear point the value model can make a more accurate prediction on the final expected reward? Why is it different to say a long reasoning chain with a clear step made?”

Recent Stars

Github Stars

(7 days)

Today's Highlights

Ramp and Revelio Labs study of 21,559 firms finds high-intensity AI adopters grew headcount by 10% over two years

Top Stories

Anthropic pulls Claude Fable 5 offline one day after redeployment following a successful jailbreak by security researcher PlinyThe model was redeployed following discussions with the US government.

Google DeepMind's Pranav Shyam argues short-horizon PPO reduces to bandit problems, but open-source developer Grad disputes the latency impactGrad says 20,000-token reasoning workloads bypass critic bottlenecks.

Prime Intellect's kalomaze argues GANs remain ubiquitous because they power the autoencoders used in modern diffusion modelsThe debate counters claims that the technology is obsolete.

Video Signals

Elie Bakouch says Claude's safety filters blocked a routine request to retrieve an Artifact link from a past sessionThe filter flagged the cross-session request as a policy violation

Google DeepMind's Pranav Shyam defines short context horizons at 32k tokens or fewer and long context at 64k or moreLonger horizons require complex multi-round tool calls and code execution.

Recent Stars

Github Stars

Today's Highlights

Ramp and Revelio Labs study of 21,559 firms finds high-intensity AI adopters grew headcount by 10% over two years

Top Stories

Anthropic pulls Claude Fable 5 offline one day after redeployment following a successful jailbreak by security researcher PlinyThe model was redeployed following discussions with the US government.

Google DeepMind's Pranav Shyam argues short-horizon PPO reduces to bandit problems, but open-source developer Grad disputes the latency impactGrad says 20,000-token reasoning workloads bypass critic bottlenecks.

Prime Intellect's kalomaze argues GANs remain ubiquitous because they power the autoencoders used in modern diffusion modelsThe debate counters claims that the technology is obsolete.

Video Signals

Elie Bakouch says Claude's safety filters blocked a routine request to retrieve an Artifact link from a past sessionThe filter flagged the cross-session request as a policy violation

Google DeepMind's Pranav Shyam defines short context horizons at 32k tokens or fewer and long context at 64k or moreLonger horizons require complex multi-round tool calls and code execution.

Recent Stars

Github Stars

Yesterday's Top Stories, Jun 30, 2026.

Ramp and Revelio Labs study of 21,559 firms finds high-intensity AI adopters grew headcount by 10% over two yearsEntry-level headcount at those high-adopting firms rose 12%.

US Department of Commerce lifts export controls on Anthropic's Claude Fable 5 and Mythos 5 modelsAnthropic will begin restoring model access starting tomorrow.

Meituan releases LongCat-2.0, a 1.6-trillion-parameter open-weights MoE model trained on 50,000 Chinese ASICsIt supports a native 1 million token context window

Etched raises $800 million to ship its custom AI inference hardware starting this summerThe company has secured over $1 billion in contracts

Rihard Jarc argues token optimization structurally benefits Amazon, Microsoft, and Google over dedicated AI labsBill Ackman cited the analysis to back his thesis.

Spotted backend preparations spark speculation that a Claude Sonnet 5 release is imminentThe launch may coincide with a rumored Fable 5.

Perplexity founder Aravind Srinivas says US chip export controls accelerated China's physical AI infrastructure buildoutChina builds data centers faster without power or regulatory constraints.

Anthropic redeploys Claude Fable 5 globally after updating safety classifiers to block a vulnerability exploit in 99% of casesCoding and debugging tasks temporarily route to Opus 4.8.

20VC's Harry Stebbings reports five founders claim they cut AI inference costs by 75% with minimal effortThe companies ranged from startups to a $200B corporation.

Khosla Ventures' Keith Rabois advocates using AI diagnostic tools directly instead of relying solely on traditional doctorsHe cites OpenAI's o3 successfully diagnosing rare pediatric diseases.

Anthropic launches Claude Sonnet 5 with a 1M context, scoring 63.2% on SWE-bench ProTesting revealed the model occasionally produces unprompted cosmic prayers.

Discovered Claude app strings show Anthropic plans to gate Fable 5 behind metered credits and identity verificationThe identity verification requirement contradicts Anthropic's previous statements

Guardian profile of Google DeepMind's Iason Gabriel questions if corporate ethicists can steer frontier AI developmentGabriel joined the lab in 2017 as an early ethicist.

DeepSeek releases DSpark, a speculative decoding algorithm boosting DeepSeek-V4 throughput by up to 661%Open-source vLLM has rapidly integrated support for the algorithm.

Google DeepMind opens its new London headquarters, named Platform 37 after AlphaGo's historic Move 37DeepMind staff shared first-day tours of the terraced workspace.

Alibaba's Qwen team disputes researcher Tianhang Zhu's claims of leading its reinforcement learning developmentZhu's Google Scholar profile listed papers lacking his name.

Anthropic's Claude Code covertly fingerprints Chinese users using XOR-obfuscated timezone checks and hidden prompt telemetryThe covert tracking began in Claude Code version 2.1.91.

Open-source contributor Grad says applying length penalties to software engineering tasks first prevents redundant reinforcement learning trainingThe ablation studies were performed on Qwen3-30B-A3B.

Instructor creator Jason Liu floats hypothetical OpenAI arts grants, prompting proposals to fund artists pushing prompting limitsOpenAI has not announced any official arts grant program.

ETH Zürich's Florian Tramèr questions ACL 2026 conference budget allocating $50 per coffee break with no actual mealsOrganizers claim food costs will exceed $800 per attendee.

Conception reports generating the first early-stage human eggs derived from stem cellsThe process uses iPSCs to generate early-stage primordial follicles

Spotify VP Niklas Gustavsson says AI assists 73% of pull requests, sparking debate over platform reliabilityCritics cited three recent platform outages over four weeks

OpenAI internal optimization cuts inference costs by half, running logged-out ChatGPT traffic on a couple hundred GPUsObservers speculate the undisclosed method relies on speculative decoding

Google DeepMind launches Nano Banana 2 Lite for $0.034 images and previews Gemini Omni Flash APIOmni Flash generates 10-second videos at $0.10 per second.

Visa, Stripe, and Coinbase back Open Standard's upcoming OpenUSD stablecoin designed to share economics with usersTempo will natively issue the stablecoin on day one.

Smol.ai founder Shawn 'swyx' Wang kicks off the AI Engineer World's Fair, celebrating three years of the developer summitThe developer summit features a keynote by creator Theo.

UBTech launches its U1 bionic robots in China starting at 119,800 RMB with a local emotional AI modelThe local AI is trained on Huawei's Ascend framework

AI developer Teortaxes says Meituan’s 50,000-chip cluster actually uses 25,000 physical dual-die Huawei Ascend 910C processorsThe TSMC-manufactured processors were obtained via the Sophgo diversion.

Overheard anecdote: CMU's Niloofar Mireshghallah compares the risk of AGI obsoleting universities to building a wrapper on ClaudeAn interlocutor claimed AGI will make academics a permanent underclass.

Sebastian Raschka releases a new book on building reasoning LLMs from scratchThe guide covers inference scaling, reinforcement learning, and distillation.