XIndustry08:06Rohan Paul@rohanpaul_aiToday’s edition of my newsletter just went out. 🔗 https://rohanpaul.substack.com/p/meta-releases-muse-spark-open-816b 🗞️ Meta Releases Muse Spark Open: 816B Open-Weight MoE With 1M Context 🗞️ Elorian releases Glimpse, a visual-thinking model that re-inspects images and video as it reasons 🗞️ After years of rivalry, Meta and Apple join hands to deliver open-weight AI to the Mac. 🗞️ Claude Mythos 5 designed an elliptic-curve quantum circuit using 17x fewer gates for 12% more qubits. 🗞️ OpenAI engineer arrested by FBI for allegedly bugging conference room 🗞️ New Google and Carnegie Mellon paper finds self-improving agents hold on to security gains far better when security is evolved as a separate goal. 🗞️ Eight states have sued OpenAI separately in their own courts, alleging ChatGPT is unsafe for children and that the company misled parents about it. 🗞️ FT first reported Anthropic’s third-quarter revenue of $16.8B grew about 45% from the second quarter, a steep drop from the more than 100% jump between the first and second quarters. 🗞️ SpaceXAI launched Grok Voice 3.0, ranks first on the Speech-to-Speech Index by Artificial Analysis 🗞️ WSJ reported how the leading AI labs are now regularly running counterespionage drills for their employees.Collected by WispRead in full↗
XIndustry03:13Rohan Paul@rohanpaul_aiNew Peking University and Samsung paper shows that a slow reasoning model can coach a fast model offline, raising its live game score from 13.51 to 24.19. shows that the stronger reasoning model in paused games can fall far behind a faster model once the game keeps moving. Many game evaluations pause the world while the model thinks, so slow answers cost nothing. In real settings, a correct action can arrive too late. They built RealtimeBench, 180 tasks across 6 games like Mario, Snake, and Tetris that keep running during inference. GPT-6.1-sol beat the faster Jev in 5 of 6 games when paused, but scored 1.51 versus 13.51 out of 100 live. Their fix, TORSI, has the slow model study the fast model's failures between games and write instructions the fast model follows during play. After 2 rounds, Jev's average score rose from 13.51 to 24.19, improving all 6 game types. For real-time agents, keep a fast model in the loop and use the slow model offline to write and test its instructions.Collected by WispRead in full↗
XIndustry02:10Rohan Paul@rohanpaul_aiUnderstanding a video means comparing moments, and models are getting much better at it. Glimpse from @ElorianAI checks how a scene changes from one point in time to another, and goes back for a closer look when the change is small or fast. Elorian calls this visual thinking: reasoning over what the model observes, grounded in the frames. Glimpse can return to the source during reasoning, crop into an area, extract specific video frames or run code to measure or compare what it sees, a mechanism Elorian calls a “visual scratch space.”Collected by WispRead in full↗
XIndustry00:33Rohan Paul@rohanpaul_aiJensen Huang explains why AI may soon need a billion times more compute. The old way was retrieval: the model pulled up what it had memorized and answered. The new way is problem solving, where the model weighs several options, rolls out the best one to three and prunes the rest. That search takes far more work than recalling a stored answer, and every step of it runs on GPUs. ---- Full video on "Bg2 Pod" YouTube channel, (link in comment)Collected by WispRead in full↗
XIndustry03:37Rohan Paul@rohanpaul_aiJensen Huang was visting Taiwan with @lexfridman . His case against export controls: build the flywheel instead of just blocking access. "So the question is, are you optimizing for someone else not having access to this technology, or are you optimizing for a flywheel? ... And I suggest that a more sophisticated long-term strategy is to optimize for a flywheel." He argues that keeping things away from rivals makes sense for physical items like refined rare earths, bullets and bombs, but he argues that for computing tech, America is safer by selling widely. Sales create wealth, wealth pays for more R&D, and that loop keeps the US ahead, because, the safest countries in the world are the richest ones. ---- Full video on "Lex Fridman" YouTube channel, (link in comment)Collected by WispRead in full↗
XIndustry02:47Rohan Paul@rohanpaul_aiMETA is narrowing the gap between its advertising business and GOOGL Search, with quarterly revenue now just about $4B short. AI enabled Meta to serve 14% more ads at 12% higher prices because it improved at predicting what people want before they even search. https://x.com/Fiscal_ai/status/2111125316873990409/video/1Collected by WispRead in full↗
XIndustry01:21Rohan Paul@rohanpaul_aiCitigroup now bars third-party AI agents from logging into customer accounts or moving money inside its digital banking platforms, according to the Financial Times. Bank of America has imposed similar limits, joining JPMorgan Chase, which is restricting outside AI agents after Jamie Dimon warned that thousands of “rogue agents” could otherwise operate inside the bank. The deeper threat is to bank funding, since agents comparing rates could move cash from checking accounts paying about 0.1% into fintech accounts paying 3.3% to 5%. JPMorgan, Bank of America and Wells Fargo hold $1.6T in non-interest deposits, and paying 3% on that money would cost $47B a year, close to half their combined annual earnings.Collected by WispRead in full↗
XIndustry00:09Rohan Paul@rohanpaul_aiThis is a big win for reproducible research. AIPOCH has released Open-Science, a free, open-source desktop app where AI agents run scientific analyses and keep the code behind every result. Their newest feature is the .science file, which packs a finished research session, including conversation branches, file versions, notebook records and verification evidence, into one file you can import on another computer. The recipient can trace how you reached each result, but the import opens read-only, so they must make a copy before running new experiments. It runs on existing agent engines such as Claude Code, OpenCode, Codex or CodeBuddy, and accepts cloud models, custom gateways, local servers like Ollama, or Claude and Codex subscriptions. What it adds is the science layer, with 23 built-in skills, 24 built-in data connectors, approval controls before risky actions, and remote jobs over SSH or Slurm.Collected by WispRead in full↗
XIndustry03:13Rohan Paul@rohanpaul_ai– https://arxiv.org/abs/2610.20665 Title: "The World Won't Wait: Evaluating and Teaching Agents Under a Live Clock"Collected by WispRead in full↗
XIndustry02:28Rohan Paul@rohanpaul_aiClaude Mythos 5 designed an elliptic-curve quantum circuit using 17x fewer gates for 12% more qubits. Andi Gu, a Harvard physics researcher, ran it through Anthropic’s STEM Fellows Program, which pairs scientists with Anthropic staff to test whether Claude can do real research in their field. Claude wrote circuit designs as code, an evaluator checked and scored them, and Claude kept revising until it beat human-designed records in most of the 8 tasks. Its strongest result made an elusive classical math shortcut work on quantum hardware for the first time, after Claude proved new identities and verified them with formal proofs.Collected by WispRead in full↗
XIndustry01:01Rohan Paul@rohanpaul_aiIt’s the first serious attempt to bring top-tier robotics into entertainment instead of factories. Hugo from Dynamic Creatures is a prototype intended as the first in a series of robots designed to interact with humans in everyday settings. Most robots register a person as an obstacle to slow down for. Hugo is built to treat your face, voice and movement as the input that drives what it does next. https://x.com/theermann/status/2111124899788210639/video/1Collected by WispRead in full↗
XIndustry22:13Rohan Paul@rohanpaul_aiBrands and studios, this one’s for your legal team. You no longer have to choose between AI and copyright. VN has launched VN Studio, an AI filmmaking workspace that records human edits as evidence for copyright claims. You generate a clip with AI, then edit it yourself, and VN’s desktop app Bridge writes each of your edits into a log, kept separate from anything the AI did. Every edit you make gets a timestamped entry that’s chained to the one before it, so nobody can quietly change your editing history later without it showing. That is exactly whats needed because copyright won’t protect what the AI made, only what you added, so this log becomes your proof of which parts are yours when you register or defend your copyright.Collected by WispRead in full↗
XIndustry12:57Rohan Paul@rohanpaul_aiJensen Hunag talks about a large revenue line of Nvidia, software licensing: charging companies a yearly license of roughly a couple of thousand dollars per GPU to run, secure and maintain Nvidia's AI software, and with about 20M GPUs shipped a year and around 50% going to enterprises, this is a huge recurring revenue stream. ---- Full video on "Bg2 Pod" YouTube channel, (link in comment)Collected by WispRead in full↗
XIndustry05:02Rohan Paul@rohanpaul_aiAfter years of rivalry, Meta and Apple join hands to deliver open-weight AI to the Mac. Meta released the weights for Muse Spark Open, an 816B-parameter model that runs on 2 linked Mac Studios. Apple tuned the release for its MLX framework. Only 42B parameters fire per token because the model routes each token to 8 of 256 experts, so every step reads a small slice of memory. Holding the whole model is the hard part, since the 4-bit NVFP4 version needs at least 700GB of GPU memory. Each M5 Ultra Mac Studio gives its GPU up to 512GB of unified memory, so 2 machines hold the model while 1.2TB/s of bandwidth feeds each token. On the benchmarks, Muse Spark Open does top the MCP-Atlas tool-use test at 89.2, which suits the coding agents Meta built it for. The license allows free commercial use but requires a “Built with Muse Spark Open” label, plus a separate Meta deal for API model sellers with over $20M in 12-month revenue. Apple hasn’t priced the 512GB Mac Studio, and a cluster of 4 machines with 256GB each costs roughly $38,000.Collected by WispRead in full↗
XIndustry15:27Rohan Paul@rohanpaul_aiThe interesting experiment is basically this: find where the 3.54 WER breaks. Run HojoAI/Hojo-ASR-Multi-V1 against the ASR stack already being used, publish the audio conditions, hardware, settings, ground truth, and raw outputs, then show exactly where each model fails.Collected by WispRead in full↗
XIndustry05:36Rohan Paul@rohanpaul_aiGoogle achieved a new level in quantum computing. Says it has cleared a bottleneck researchers have flagged since 2004. Quantum computers need single-use “magic states” to run the operations that put them beyond ordinary computers. Making clean ones is so costly that some estimates say it could take up to 95% of a machine’s capacity. Google’s new method checks each state and throws out flawed attempts, cutting errors about 40x on its 105-qubit Willow chip. The surviving states reached 99.99% accuracy, a record by Google’s count, in work published in Science today. Only 8% of attempts survived the filter, and the method still has to work at much larger scale. Google has shown on real hardware 1 of the methods behind its estimate that RSA-2048 encryption could be broken with under 1M qubits. https://x.com/sundarpichai/status/2110804295352488143/video/1Collected by WispRead in full↗
XIndustry04:54Rohan Paul@rohanpaul_aiMost AI image failures come from the model inventing everything at once: hands, light, composition. "Edit with AI" on Envato starts from a human-shot stock photo and changes only what you select or describe, so the real structure survives. This is definitely a much more practical way to create better final result. i.e. a real good photo has already settled the subject, framing and light, so the AI only handles the change you asked forCollected by WispRead in full↗
XIndustry15:27Rohan Paul@rohanpaul_aiAn open multilingual ASR model just posted 3.54 average WER across 5 languages Hojo-ASR-Multi-V1 combines an acoustic encoder with a Qwen3-4B decoder, giving speech recognition an unusually strong language model at the final decoding stage. - Built with a Qwen3-4B LLM decoder - Designed for multilingual conversational speech recognition - Publicly available on Hugging Face, under the Apache 2.0 license - Installable through the hojo-asr Python package - Supports local inference - Accepts WAV file paths, audio byte streams and SCP inputs There is a bigger technical story behind the 3.54 WER number. Traditional ASR systems rely heavily on acoustic and language modeling components designed specifically for speech, while Hojo hands the decoding stage to a Qwen3-4B LLM after the audio has been encoded and adapted. So recognition gets access to a much richer model of how words naturally fit together while it is deciding what was actually said. What makes Hojo even more interesting is that it is reaching that level across multiple languages rather than being built around only 1 language. And it does this with an openly available model that developers can actually run and inspect themselves. The architecture is probably the most interesting part for me. Hojo uses an Encoder, Adapter and LLM setup with Qwen3-4B as the decoder, so the final transcript is produced using both acoustic information from the speech and the language knowledge inside the LLM. That gives the decoder much more context for resolving words that may sound ambiguous from audio alone.Collected by WispRead in full↗
XIndustry05:13Rohan Paul@rohanpaul_aiNew York City’s comptroller asked Anthropic to delay its IPO until safety concerns are resolved. He also oversees more than $325B in New York City pension assets and has an exposure to Anthropic of about $7M. He wants Anthropic to detail its planned audits and safeguards, the risks they address and an implementation timeline, and he calls a “morally binding” White House accord insufficient.Collected by WispRead in full↗
XIndustry04:41Rohan Paul@rohanpaul_aiMeta released the weights of Muse Spark Open, on Huggingface, an 816B-parameter model built for coding and agentic work. The open version shares the same base model as Muse Spark 1.3, with additional safety post-training to resist misuse as a standalone model without system guardrails, reversing Meta’s first proprietary, closed-weight frontier model strategy from April. Each of its 62 layers holds 256 expert networks and routes every token through 8, so deployment needs the memory of an 816B-parameter model but runs at the speed of a 42B model. That storage bill explains the Apple tie-up, because the M5 Ultra pushes 1.2 TB/s of unified memory and a Mac Studio can carry up to 512GB. A 3-bit quant tested on a 512GB M3 Ultra took about 344 GB and ran at about 21 tok/s generation on a short prompt.Collected by WispRead in full↗