ニュース

This is today’s edition of The Download, our weekday newsletter that provides a daily dose of what’s going on in the world of technology. Anthropic found a hidden space where Claude puzzles over concepts The AI firm Anthropic has got the clearest glimpse yet at what’s really going on inside large language models as they answer questions or carry out tasks. What they found ranges from the mundane to the unnerving.  Researchers at the company built a tool called the Jacobian lens (or J-lens) and used it to uncover a hidden area, which they named the J-space, inside its flagship LLM, Claude. The J-space contains words related to the response a model is working on but may not ultimately produce. If Claude were a person (which it is not), you might say these hidden words reveal what’s on its mind before it actually speaks.  Read the full story on what they found. —Will Douglas Heaven The must-reads I’ve combed the internet to find you today’s most fun/important/scary/fascinating stories about technology. 1 OpenAI has unveiled its long-awaited “super app” ChatGPT Work blends its chatbot, coding tool, and new models. (Reuters $)+ It’s designed to do your work for you and with you. (Ars Technica)+ And arrived the same day as OpenAI’s GPT 5.6 models. (NYT $)+ It’s also developing a fully automated researcher. (MIT Technology Review) 2  Humanoids have performed teleoperated surgery on living animalsIn the world-first, they removed gallbladders from pigs. (Ars Technica)+ The human work behind humanoids is hidden. (MIT Technology Review) 3 SK Hynix has landed the largest US listing by a foreign companyThe South Korean chip giant raised $26.5 billion. (CNN)+ Demand for AI data centres has led its profits to skyrocket. (Guardian)+ But its jumbo share sale may be a sign of overheated times. (FT $)+ South Korea’s hottest bachelors are chip workers. (MIT Technology Review) 4 Tencent is leading a deal to unwind Meta’s $2 billion Manus acquisitionIt’s in talks to become the Chinese AI startup’s largest shareholder. (FT $)+ Tencent will reportedly buy Manus for no less ​than $2 billion. (Reuters $)+ Beijing had ordered Meta to unwind the acquisition. (Bloomberg $) 5 Resuscitated human retinas responded to light 10 hours after deathIt’s a big step towards eye transplants that restore vision. (New Scientist $)+ As is a new device that revives dead eyeballs. (MIT Technology Review) 6 Meta has started charging for AI accessA new version of Muse Spark has a paid tier for developers. (Quartz) + Meta also plans to start producing an AI chip in September. (Reuters $) 7 OpenAI and Google have sold AI models to blacklisted China groupsVia Singapore-based subsidiaries of Alibaba, Baidu and Tencent. (FT $) 8 A daughter tested an AI “death bot” of her fatherThe technology provided both comfort and unease. (New Yorker $) 9 An astronomer says the hunt for alien life needs more statisticsHe wants to replace speculation with mathematical frameworks. (Quanta) 10 Pokémon Go players turned Times Square into a giant battlefieldMore than 1,500 fans finally fulfilled the game’s 2016 launch promise. (Wired $)+ Pokémon Go is also training world models. (MIT Technology Review) Quote of the day “When we’re talking about AI, we love the hype, we get excited about it. The damn thing never actually lands in practice.” —Vijay Janapa Reddi, an engineering professor at Harvard University, tells Wired why he’s skeptical about grand plans for AI. One More Thing B.F. SKINNER FOUNDATION Why we should thank pigeons for our AI breakthroughs In 1943, psychologist B.F. Skinner led a secret government project to make bombs more precise. His idea: teach pigeons to guide missiles by pecking at targets on a screen inside a warhead. To train them, Skinner rewarded the birds with food when they made the right decisions, using trial and error to shape their behavior. Unsurprisingly, the military never deployed Skinner’s kamikaze pigeons. Yet his experiments convinced him that pigeons were “an extremely reliable instrument” for studying learning.   Decades later, those same principles would help power reinforcement learning, the technology behind some of today’s most advanced AI systems. Discover how pigeons inspired one of AI’s most powerful techniques. —Ben Crair We can still have nice things A place for comfort, fun, and distraction to brighten up your day. (Got any ideas? Drop me a line.) + Here’s a splendid selection of this year’s NSW architecture award winners.+ Photographers have captured the Strawberry Moon’s golden glow in stunning detail.+ Idiocracy is the film that best exemplifies the “American experience,” according to a new poll. Look back at the prescient comedy with this Screen Junkies trailer.+ Get ready for the weekend with this psychedelic house journey from Jamie xx b2b Caribou.

Large hybrid MoE models like Nemotron-3-Super are accurate but expensive to serve. Their active parameters, KV cache, and Mamba state cap how many users a node can hold at a given per-user token rate. NVIDIA AI team has released Nemotron-Labs-3-Puzzle-75B-A9B, a compressed variant of Nemotron-3-Super. The parent model has 120.7B total and 12.8B active parameters. The compressed model has 75.3B total and 9.3B active parameters. The deployment target was fixed before the architecture search began. Target one was 2x server throughput at 100 tokens per second per user. Target two was 8 concurrent 1M-token requests on a single H100. Three checkpoints on Hugging Face: BF16, FP8, and NVFP4. TL;DR 120.7B/12.8B active compresses to 75.3B/9.3B active, with the 88-block hybrid layout preserved. 8xB200 total throughput rises 1.60x to 2.14x over Super at matched NVFP4 and matched user throughput. Single-H100 1M-token concurrency goes 1 to 8, driven by a 70 GB to 44.5 GB weight drop. Iterative Puzzle beats single-step Puzzle by 0.57 average points at the same compression target. Arena-Hard-V2 (-4.2) and SWE-Bench (-2.6) are the real costs; RULER and AA-LCR barely move. Nemotron-Labs-3-Puzzle-75B-A9B Nemotron-3-Super is a hybrid Mamba-Transformer MoE model. Puzzle-75B-A9B preserves the parent’s block layout exactly. It has 88 blocks: 40 Mamba, 40 MoE, and 8 attention blocks. What changed is capacity inside those blocks: Quantity Super Puzzle-75B-A9B Ratio Total parameters 120.7B 75.3B 62.4% Active parameters 12.8B 9.3B 73.1% Mamba SSM state size 128 96 75% MoE routed expert intermediate size 2688 1280-2688 Mean 59.9% Activated routed experts per token 22 4-18 Mean 50% Active routed expert capacity (relative) 100% 8.7%-62.3% Mean 30.9% The number of routed experts, the shared expert size, and the MoE latent size are unchanged. Attention layers were left untouched. The proposed research’s stated reason is that Nemotron-3-Super is already very KV-cache efficient. Mamba layers were pruned uniformly, because inference frameworks do not support a different SSM state size per layer. https://arxiv.org/pdf/2607.04371 The result is not a uniformly scaled-down teacher. The above figure shows the allocation across depth. Puzzle preserved capacity in selected middle and late layers, and cut hard elsewhere. Benchmark and Performance The below table reports Pareto-optimal total throughput on a single 8xB200 node, with single-step decoding. Scenario (in/out) UT floor Super (tok/s) Puzzle-75B-A9B (tok/s) Boost 50K / 2K >= 100 5,128 8,210 1.60x 50K / 2K >= 125 3,784 6,412 1.69x 50K / 2K >= 150 2,532 4,523 1.79x 8K / 64K >= 100 20,939 42,601 2.03x 8K / 64K >= 125 13,074 27,918 2.14x 8K / 64K >= 150 8,522 18,047 2.12x Both models were served at matched NVFP4 weights, FP8 KV cache, and FP16 Mamba state. The gap therefore reflects compression, not a change in numeric format. The prefill-heavy 50K/2K regime gains least. The decode-heavy 8K/64K regime gains most. On a single 8xH100 node at UT = 100, the gains are smaller. They are 1.91x on 50K/2K and 1.82x on 8K/64K. Both models there use FP8 weights, FP8 KV cache, and FP32 Mamba state. On a single H100 at 1M context, the binding constraint flips from compute to memory. Super’s NVFP4 weights occupy about 70 GB of the 80 GB HBM budget. Each 1M-token request adds about 4 GB of KV cache. Effective concurrency is therefore 1. Puzzle-75B-A9B’s NVFP4 weights occupy about 44.5 GB. Attention layout is unchanged, so per-request KV cost is unchanged. Concurrency at 1M rises to 8. Aggregate decode throughput at that concurrency is roughly 4x Super’s single-request throughput. Prefill of a 990K-token prompt is about 1.2x faster. How Iterative Puzzle Works Puzzle is a decomposed neural architecture search framework, implemented here as Puzzletron. It defines a discrete search space of alternative layer implementations. Each alternative gets a quality score. A mixed-integer program then selects one alternative per layer under a deployment constraint. Three pruning techniques form the search space: Intermediate channel pruning: Channels inside each routed expert are ranked by contribution to the expert’s output. All experts within one MoE layer are pruned to a uniform size, for kernel compatibility. Top-k reduction: The number of experts a token is routed to varies per layer, up to the parent’s k=22. Mamba SSM pruning: The SSM state size drops from 128 to 96 channels. The SSM result is measured. Dropping 128 channels to 96 speeds the SSM kernel 1.2x to 1.3x during decode. This holds at batch sizes between 8 and 512. Channels were ranked by estimated contribution to the Mamba layer output. The estimate averaged over 67M tokens of validation data. Appendix A shows this beats random channel selection under aggressive pruning. The original formulation assumes replacement quality impacts are approximately additive. Each candidate block is scored inside the unmodified parent. That ignores higher-order interactions between replacements. Iterative Puzzle alternates bounded compression with short knowledge distillation recovery. It builds a sequence M0, M1, … MR instead of jumping to the target. Scores are recomputed against the current compressed model, not the original parent. Three stages were used: MoE weights to 75% of teacher capacity, Mamba SSM state to 75%. Healed for 24B tokens. MoE weights to 60% of teacher capacity. Healed for 43.2B tokens. Activated routed-expert budget to 50%, allocated heterogeneously. Healed for 52.8B tokens. https://arxiv.org/pdf/2607.04371 The above table compares this against a single-step Puzzle baseline at the same target. The three-step procedure averages 69.05 across ten benchmarks, against 68.48. Gains appear on MMLU-Pro, GPQA, HLE, AA-LCR, LiveCodeBench, SciCode, and RULER-256K. IFBench-Instruction fell 0.2 points and IFBench-Prompt fell 0.5. Recovery: Distillation, RL, and Verbosity Knowledge distillation ran on 30% pretraining data and 70% SFT data from Nemotron-3-Nano. During the Puzzle phase, KD used a 32K sequence length. Recovery then trained at 128K, and scaled to 512K. The budget was up to 100B tokens, with a 16M-token global batch, in Megatron-LM. RL post-training adopted Stage 2 of the Nemotron-3-Super RL pipeline, focused on software engineering. Phase 2.1 did single-step tool-use comparison. Phase 2.2 moved to end-to-end sandbox RL, where agents run up to 200 turns. Both phases used a KL penalty of 0. The team swept learning

The default assumption in most LLM developer communities is that you start with raw API calls and graduate to a framework as your project grows.

This is today’s edition of The Download, our weekday newsletter that provides a daily dose of what’s going on in the world of technology. Four nuclear reactors hit a big milestone in the US —Casey Crownhart I was really looking forward to July 4, and not just because I love a poolside barbecue. This year the American holiday also marked a big symbolic deadline for US nuclear power. Last year the Trump administration set a goal to see three new microreactors achieve criticality, a technical milestone establishing that a reactor can sustain a chain reaction, by the nation’s 250th birthday. And just in time, not just three, but four reactors did so. It’s a positive sign for nuclear technologies at a time of increasing need for electricity and emissions-free energy sources. But achieving criticality doesn’t mean a reactor is ready to provide electricity for the grid (or at all, for that matter). Here’s what the milestone could mean for nuclear power in the US—and where the four companies might go next. This story is from The Spark, our weekly climate tech newsletter. Sign up to receive it in your inbox every Wednesday. The must-reads I’ve combed the internet to find you today’s most fun/important/scary/fascinating stories about technology. 1 China plans to let its top AI firms buy Nvidia H200 chipsAlibaba, ByteDance, and DeepSeek are set to get permission. (Information $)+ China had previously withheld approval despite US authorization. (Reuters $) 2 NATO is building a network to stop Russian attackers in their tracksIt will use sensors, drones, satellites, and AI to detect them. (Business Insider)+ Troops are donning odd camouflage to elude drones. (Economist $)+ The US wants cheaper drones as Iran’s wrecking its Reapers. (Ars Technica) 3 Researchers have a new idea to fight future El Niños: dimming the sunDeflecting solar energy could cool the ocean and mitigate the risks. (Wired $)+ But there could be unexpected consequences. (New Scientist $)+ And geoengineering as a field is getting a reality check. (MIT Technology Review) 4 Meta is patenting an AI device that records users to analyse emotionsIt ostensibly aims to tailor workout plans to the user’s mood. (404 Media)+ AI memory is privacy’s next frontier. (MIT Technology Review) 5 Chipmakers are going vertical as Moore’s Law slowsThey’re stacking transistors to keep chips advancing. (Economist $)+ IBM is betting on the technique. (MIT Technology Review) 6 Ivy League students suspected of AI cheating saw scores fall in personFrom 96% all the way down to 48%. (Ars Technica)+ AI giants want to take over the classroom. (MIT Technology Review) 7 A new study says parents’ phone addictions damage bonds with kidsIt can exacerbate “insecure attachment” for life. (Bloomberg $)+ And make children more anxious and avoidant. (Gizmodo) 8 A judge approved Musk’s $1.5 million Twitter settlement with the SECDespite what she called “serious misgivings” and “red flags.” (Reuters $)+ Musk was accused of skirting stock disclosure rules. (Fortune) 9 Shoebox-sized “detector satellites” could find nuclear bombs in spaceCubesats carrying the detector could sense a bomb’s radiation. (Space)+ Russia is suspected of developing space-based nukes. (Reuters $) 10 A World Cup match drove Google Search traffic to a new recordThe milestone came after Argentina’s comeback against Egypt. (CNBC) Quote of the day “I talk about it on Tic Tac.” —President Donald Trump tells the public where to find his insights on the dangers of communism, Gizmodo reports. One More Thing Robots are bringing new life to extinct species Paleontologists aren’t easily deterred by evolutionary dead ends or a sparse fossil record. And in the last few years, they’ve developed a new trick for turning back time and studying prehistoric animals: building experimental robotic models of them.  In the absence of a living specimen, an ambling, flying, swimming, or slithering automaton is the next best thing for studying the behavior of extinct organisms. Learning more about how they moved can in turn shed light on their lives, such as their historic ranges and feeding habits. Scientists can simply sit back and observe their behavior in different environments.  Read the full story on the rise of paleo-inspired robots—and four examples that are shedding light on creatures of yore. —Shi En Kim We can still have nice things A place for comfort, fun, and distraction to brighten up your day. (Got any ideas? Drop me a line.) + Georgia Hill’s monochrome artworks are filled with visual harmony.+ AI has salvaged text from a papyrus scroll burned to a crisp when Mount Vesuvius erupted 2,000 years ago.+ Rare images taken by a Japanese space probe show a near-Earth asteroid resembling a cuddly snowman.+ “Another One Bites the Bee Gees” smoothly merges two classic tracks with a 4/4 time signature into the perfect song for applying CPR.

This is today’s edition of The Download, our weekday newsletter that provides a daily dose of what’s going on in the world of technology. Four nuclear reactors hit a big milestone in the US —Casey Crownhart I was really looking forward to July 4, and not just because I love a poolside barbecue. This year the American holiday also marked a big symbolic deadline for US nuclear power. Last year the Trump administration set a goal to see three new microreactors achieve criticality, a technical milestone establishing that a reactor can sustain a chain reaction, by the nation’s 250th birthday. And just in time, not just three, but four reactors did so. It’s a positive sign for nuclear technologies at a time of increasing need for electricity and emissions-free energy sources. But achieving criticality doesn’t mean a reactor is ready to provide electricity for the grid (or at all, for that matter). Here’s what the milestone could mean for nuclear power in the US—and where the four companies might go next. This story is from The Spark, our weekly climate tech newsletter. Sign up to receive it in your inbox every Wednesday. The must-reads I’ve combed the internet to find you today’s most fun/important/scary/fascinating stories about technology. 1 China plans to let its top AI firms buy Nvidia H200 chipsAlibaba, ByteDance, and DeepSeek are set to get permission. (Information $)+ China had previously withheld approval despite US authorization. (Reuters $) 2 NATO is building a network to stop Russian attackers in their tracksIt will use sensors, drones, satellites, and AI to detect them. (Business Insider)+ Troops are donning odd camouflage to elude drones. (Economist $)+ The US wants cheaper drones as Iran’s wrecking its Reapers. (Ars Technica) 3 Researchers have a new idea to fight future El Niños: dimming the sunDeflecting solar energy could cool the ocean and mitigate the risks. (Wired $)+ But there could be unexpected consequences. (New Scientist $)+ And geoengineering as a field is getting a reality check. (MIT Technology Review) 4 Meta is patenting an AI device that records users to analyse emotionsIt ostensibly aims to tailor workout plans to the user’s mood. (404 Media)+ AI memory is privacy’s next frontier. (MIT Technology Review) 5 Chipmakers are going vertical as Moore’s Law slowsThey’re stacking transistors to keep chips advancing. (Economist $)+ IBM is betting on the technique. (MIT Technology Review) 6 Ivy League students suspected of AI cheating saw scores fall in personFrom 96% all the way down to 48%. (Ars Technica)+ AI giants want to take over the classroom. (MIT Technology Review) 7 A new study says parents’ phone addictions damage bonds with kidsIt can exacerbate “insecure attachment” for life. (Bloomberg $)+ And make children more anxious and avoidant. (Gizmodo) 8 A judge approved Musk’s $1.5 million Twitter settlement with the SECDespite what she called “serious misgivings” and “red flags.” (Reuters $)+ Musk was accused of skirting stock disclosure rules. (Fortune) 9 Shoebox-sized “detector satellites” could find nuclear bombs in spaceCubesats carrying the detector could sense a bomb’s radiation. (Space)+ Russia is suspected of developing space-based nukes. (Reuters $) 10 A World Cup match drove Google Search traffic to a new recordThe milestone came after Argentina’s comeback against Egypt. (CNBC) Quote of the day “I talk about it on Tic Tac.” —President Donald Trump tells the public where to find his insights on the dangers of communism, Gizmodo reports. One More Thing Robots are bringing new life to extinct species Paleontologists aren’t easily deterred by evolutionary dead ends or a sparse fossil record. And in the last few years, they’ve developed a new trick for turning back time and studying prehistoric animals: building experimental robotic models of them.  In the absence of a living specimen, an ambling, flying, swimming, or slithering automaton is the next best thing for studying the behavior of extinct organisms. Learning more about how they moved can in turn shed light on their lives, such as their historic ranges and feeding habits. Scientists can simply sit back and observe their behavior in different environments.  Read the full story on the rise of paleo-inspired robots—and four examples that are shedding light on creatures of yore. —Shi En Kim We can still have nice things A place for comfort, fun, and distraction to brighten up your day. (Got any ideas? Drop me a line.) + Georgia Hill’s monochrome artworks are filled with visual harmony.+ AI has salvaged text from a papyrus scroll burned to a crisp when Mount Vesuvius erupted 2,000 years ago.+ Rare images taken by a Japanese space probe show a near-Earth asteroid resembling a cuddly snowman.+ “Another One Bites the Bee Gees” smoothly merges two classic tracks with a 4/4 time signature into the perfect song for applying CPR.