The Archive

Temperature-zero BF16 LLM inference is often treated as reproducible, yet the same request can emit different tokens when decoded alone or inside a larger batch. Existing fixes use batch-invariant operators or LLM-42's per-token verification, incurring cost even when most steps are stable. We ask whether verification can be applied exclusively to flipped tokens. Across five models, batch-induced token flips are sparse on the flip-rate benchmarks: on MATH500, Llama-3.1-8B flips on $0.48\%$ of synchronous decode steps, and all tested models stay within the 0.3-1.3% range on MATH500, GSM8K, and ...

I was at 88% last night and woke up until 4pm to optimize my agents so I can work during the weekend. But after waking up, my usage is all 0 now, I checked in the app, on the web, all showing zero. Did AI God grant me a wish? Edit: wow Opus 4.8 is here, AI God really grant us all a wish

Third-person singular pronouns have long been used to study stereotypical biases in language models and to test their abilities to reason about reference. More recently, the interplay between reasoning and bias has been investigated with the task of pronoun fidelity, which assesses models' abilities to correctly reuse a previously-specified pronoun for a discourse entity, independent of other potentially distracting discourse entities mentioned in between. However, such research focuses on English, which is a language with limited grammatical gender and almost no gender agreement. In this pap...

Opus 4.8 is live

Continuous-time models are a natural choice for irregular and asynchronous data. A central design choice is how to embed discrete observations into continuous time. Interpolation- and imputation-based embeddings reconstruct a continuous observation path, making the model sensitive to the choice of reconstruction. We show that this reconstruction step is unnecessary; under mild conditions, compact-set universality on the model input space transfers to the data space whenever the embedding from data to input is continuous and injective. Guided by this result, and building on the rectilinear con...

https://preview.redd.it/ijwlm2f2pw3h1.png?width=2536&format=png&auto=webp&s=9ed960f06a4f3f077d05a8557059e5534b2d1ab5 It looks like the new CC release will have opus 4.8 1M to be released anytime! I wonder if it is based of of mythos?

looks like you can run it on any potato (A1B)! [https://huggingface.co/LiquidAI/LFM2.5-8B-A1B-GGUF](https://huggingface.co/LiquidAI/LFM2.5-8B-A1B-GGUF) from LiquidAI: LFM2.5 is a new family of hybrid models designed for on-device deployment. It builds on the LFM2 architecture with extended pre-training and reinforcement learning. * **On-device personal assistant**: Designed to power real-life applications, chaining tool calls, and following complex instructions on all devices. * **Compressed performance**: Competitive with much larger dense and MoE models on instruction following and agen...

Next month's Tribeca Festival will include the premiere of an AI-generated film: Dreams of Violets. The 75-minute film is a fictional dramatization of the Iranian government's mass killing of protestors in January, with the people and images fully created by AI, as reported earlier by The Hollywood Reporter. Dreams of Violets cost $2,000 to make and is "based on journalistic reports, photographs, and eyewitness accounts," according to a press release. It was created by Ash and Pooya Koosha, two brothers who left Iran in 2009. Pooya co-founded Fountain 0, the company behind the film, while Ash...

New features coming to YouTube could make it better for listening to podcasts, rolling out to Premium subscribers starting today on Android and coming later to iOS. A new "on-the-go mode" shifts YouTube into an audio-first layout, with larger, simplified playback buttons, a still image in place of the video, and a timeline showing video chapters. YouTube says you can turn on this new mode in a video's settings - a pop up will also appear if YouTube detects you're moving around while watching a video. If you like to speed up your podcasts to get through episodes faster, YouTube's new auto spee...

Elon Musk is publicly reframing xAI’s massive Anthropic compute deal as short-term and cancellable, despite SpaceX’s own S-1 filing describing payments through May 2029.

Sesame’s new iOS app brings its conversational AI agents to the public, offering more natural back-and-forth interactions designed to feel less like traditional chatbots and more like talking to a person.

Going through the credit card statement, here's what I had active: Claude Pro (40), ChatGPT Plus (20), Cursor (20), Perplexity Pro (20), Notion AI (10), Granola (20), ElevenLabs Starter (5), Midjourney Basic (10), Gamma Pro (10), Beautiful.ai (12), Otter Pro (17), Loom Business (15), Zapier Pro (30), Make Core (10), Tactiq Pro (8), Descript Creator (15), Reclaim.ai Pro (8), Motion (19), Superhuman (30), one i can't remember the name of (10), some ai-something for instagram captions (11) Then I sat down and wrote next to each one the last time I'd actually used it. Not opened it, used it for...

Here are 12 of the biggest Google I/O 2026 keynote moments, including news about Gemini Omni, Gemini 3.5 Flash and more.

Apple's long-awaited Siri overhaul, expected to arrive in iOS 27, might look a lot like ChatGPT with a splash of Liquid Glass. Renders from Bloomberg offer a preview of iOS 27, including the new app and chat interface for Siri. The renders are "based on information viewed by Bloomberg and people with knowledge of [Apple's] plans," and could differ from Apple's final designs, which Bloomberg's Mark Gurman says Apple will reveal at WWDC in June. The images show a new pill-shaped Siri chat bubble popping out of the Dynamic Island with a drop-down menu containing options for Ask, Siri, and ChatGP...

Niche use case but I cook a lot and I've been trying to use AI tools for it consistently. Honest writeup. Works: Asking for substitutions when I'm missing an ingredient. Reliable. Tells me what to swap and why. Scaling recipes up or down with non-trivial math (recipe serves 4, I need 7 servings, what are the new quantities). Faster than I'd do it myself. Cleaning up a recipe from a website where the actual instructions are buried under 4,000 words of SEO content. Paste the URL or text, get just the recipe. Worth it for this alone. Building shopping lists from a week of planned recipes. C...

A new crop of AI labs are focused on recursive self-improvement — but the goal is proving elusive.

Enterprise AI is entering a different phase now, one where enterprises are no longer evaluating whether AI is exciting. They are evaluating whether it is safe to deploy broadly.

The update signals YouTube's ongoing efforts to compete with other platforms for podcast audiences.

Hi! Andi from Hugging Face here! My team has been working over the last few months on creating a super smooth local experience for conversations with Reachy Mini, see the video! We hope people can extend this into tons of different cool use-cases. We wrote a blog explaining how to set this up, and how to modify it for tons of different use cases. Even if you don't have a Reachy Mini, you can use this as a roadmap for amazing voice agents: [https://huggingface.co/blog/local-reachy-mini-conversation](https://huggingface.co/blog/local-reachy-mini-conversation) Hope you enjoy it!

CNN has filed a lawsuit against Perplexity, claiming that the startup's AI tools generate "verbatim" copies of its work, as reported earlier by CNN. The lawsuit, filed in a New York court on Thursday, also alleges that Perplexity provides users with information locked behind CNN's subscription. Perplexity, which offers an AI "answer" engine along with the AI browser Comet, is accused of ignoring CNN's efforts "to recognize or block Perplexity's unidentified crawlers" from scraping its content. "Human beings report, research, write, edit, and create the content that Perplexity takes without pe...

Visa said that over 1,000 employees has been using Replit for prototyping and development

Today, I’m talking with Wassym Bensaid, the chief software officer at Rivian, and the co-CEO of Rivian’s platform joint venture with Volkswagen, which everyone just calls RV Tech. That joint venture kicked off about a year and a half ago with a nearly $6 billion investment from Volkswagen. It effectively puts Wassym in charge of the operating system and electrical architecture for every future EV from Volkswagen and its associated brands, including familiar names like Audi, but also new companies like Scout. There’s a lot of Decoder ideas in there — I really wanted to know how that joint vent...

Been following the infrastructure side of AI more lately and stumbled on this from Zai. They upgraded the network architecture on a thousand-GPU cluster running GLM-5.1 coding inference from the standard ROFT setup to something they built called ZCube, developed with Tsinghua University and HarnetsAI The numbers from production: \- Switch and optical module costs down 33% \- GPU inference throughput up 15% \- P99 tail latency on first token dropped 40.6% Same GPUs, same software stack, same model. Just the network architecture changed The actual problem they were solving is interesting....

Imagine a world run by AI agents. What does it look like? What are the values or societal priorities? Is it a safer or more dangerous world? Enterprise AI startup Emergence AI is trying to find out. The company just launched Emergence World, a research lab dedicated to stress-testing the long-term viability of continuously-running AI systems. The organization ran five 15-day simulations, each governed by a different AI: Claude, ChatGPT, Grok, Gemini, and a fifth simulation run by a mix of models to see what kind of world each one builds, and whether it holds. Each simulation netted wildly d...

General Compute is betting SambaNova will be the next breakout chipmaker.

Hello everyone. The new dataset is named MONET, is Apache 2.0 and available on HF: [https://huggingface.co/datasets/jasperai/monet](https://huggingface.co/datasets/jasperai/monet) **MONET is open, Apache 2.0-licensed image–text dataset. It was built from 2.9 billion images and refined to 104.9 million high-quality samples.** We are also publishing [a paper](https://arxiv.org/abs/2605.21272) that explains how the dataset was created if you are curious and 3 compagnions projects * [A umap to visualize the distribution](https://huggingface.co/spaces/jasperai/monet-umap) * [A retreival tool ...

a feature that was requested a lot: [https://huggingface.co/models?base\_model\_relation=base](https://huggingface.co/models?base_model_relation=base)

Learn how Endava uses Codex to build an agentic organization, accelerating software delivery and reducing requirements analysis from weeks to hours.