Committee

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. A startup claims it broke through a bottleneck that’s holding back LLMs AI startup Subquadratic came out of stealth last month with a huge claim: it had solved a mathematical bottleneck that had held back large language models for almost a decade. The purported breakthrough comes from slashing the number of computations transformers need to carry out to generate answers. The result is a faster and cheaper LLM that uses far less energy than any other model on the market. Many experts remained skeptical—but Subquadratic has started to share the receipts. They suggest that their approach might be worth paying attention to. Here’s how the system works—and why some researchers still aren’t convinced. —Will Douglas Heaven Brain-computer interface trials are taking off —Jessica Hamzelou This week, I covered the story of Casey Harrell—a man with ALS who is “the first power user” of a brain implant. The device has enabled him to maintain an income, reconnect with friends and family, and read to his daughter. He told me that it’s “nothing short of revolutionary.”  Over the past couple of years, the number of BCI trial volunteers has soared. This year, China became the first country to approve a BCI for medical use. Advances in technology are allowing engineers to provide more features than ever. BCI research is properly taking off. Find out how the technology is edging from the lab towards the market. This story is from The Checkup, our weekly newsletter giving you the inside track on all things biotech. Sign up to receive it in your inbox every Thursday. The must-reads I’ve combed the internet to find you today’s most fun/important/scary/fascinating stories about technology. 1 Amazon workers who backed data center limits may face terminationThe engineers say they’re under investigation by the company. (NYT $)+ And could face discipline, including potential termination. (The Verge)+They had testified at meetings about pausing data centers. (CNBC)+ They’ve filed a joint complaint to Seattle’s Office for Civil Rights. (Wired $) 2 A new fossil discovery has rewritten 150 years of evolutionary theoryIt suggests early land vertebrates skipped the tadpole stage. (New Scientist $)+ And raises questions about how vertebrates adapted to land. (404 Media)+ Sponges may have been the first animals. (MIT Technology Review) 3 Bernie Sanders plans to give the public direct ownership of AI firmsHe’s unveiled new legislation to create an AI sovereign wealth fund. (AP News)+ It would be funded through a one-time tax on AI companies’ stock. (Quartz)+ And make annual payments directly to Americans. (Washington Post $)  4 Investors in China secretly acquired stakes in SpaceX before its IPOOne had ties to Chinese military contractors. (ProPublica)+ The US fears China has got one of ASML’s top machines. (Reuters $) 5 Researchers have figured out Russia’s nuclear-powered missileThey call it “a terrible idea”—but not an impossible one. (NPR)+ NASA is building a nuclear reactor-powered spacecraft. (MIT Technology Review) 6 Longevity medicine faces a do-or-die moment in a landmark trialIt will test whether cellular aging can be safely reversed in humans. (Axios)+ The next step is “chemical reprogramming.” (MIT Technology Review) 7 Studies suggest AI may already be deskilling professionalsOver-reliance appears to weaken doctors’ and engineers’ abilities. (Nature) 8 Tech workers who maxed out their AI use are now trying to minimize itSpiralling costs mean “tokenminning” has replaced “tokenmaxxing.” (NYT $) 9 Scientists say the human genome’s structure may confound AI modelsWhich would constrain AI-based models of biology and disease. (Quanta) 10 A new robotic self-driving toilet brings the bathroom to youThe Xiaoban also cleans up and empties itself all on its own. (The Verge) Quote of the day “They hated me. They were doing everything they could to knock me down. And look at them now.”  —Donald Trump mocks Mark Zuckerberg and Jeff Bezos in a conversation with Elon Musk that’s recounted in a new book, Wired reports.  One More Thing PABLO DELCAN Technology can help us feed the world, if we look beyond profit The pandemic exposed the weak spots in our interconnected food system. They’re the result of decades’ worth of technological advances, from globe-spanning shipping to refrigeration networks. But technology is not inherently opposed to sustainable and resilient food systems. Powerful technologies like genetic modification can create stronger local agriculture and a healthier food system—but they normally aren’t. The challenge is ensuring they serve food security and human well-being, rather than simply maximizing profits. Dive into our food system’s problems and the solutions that technology can provide. —Fabio Parasecoli 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.) + This intriguing video tracks the covert reality of Japan’s shinobi.+ Dive into this admirably obsessive archive covering over 100 different ways to tie your shoes.+ One of the world’s largest digital collections of plants and fungi is now available for free to everyone.+ A grand orchestra has beautifully covered Michael Jackson’s “Human Nature” at Abbey Road Studios.

Underneath an Apennine massif, below the Jinping Mountains of Sichuan, and at the bottom of a South Dakota mine, there is a cosmic hunt afoot. Isolated deep beneath these rocky shields, massive detectors filled with liquid xenon aim to make the first direct detections of dark matter, the long-sought invisible substance whose gravity has sculpted our universe. The hope is that someday, a bit of dark matter called a weakly interacting massive particle (a WIMP, for short) will collide with a xenon atom, creating a burst of light and electric charge. After running for years, these experiments have recently begun seeing infrequent blips from a particle that glides ethereally through ordinary matter until it crashes into the detectors. Unfortunately, the new signal is not produced by dark matter. Instead, the detectors are picking up on something similarly insubstantial but much more mundane: neutrinos, the featherweight subatomic particles that the sun and other stars produce in massive quantities. Physicists’ failure to find dark matter where they thought it was has led to a cornucopia of proposals for new ways to search: quantum sensors, liquid-helium-based detectors, searches in Jupiter’s atmosphere, and more. Physicists have known for decades that this neutrino background was there; they were just hoping to discover WIMP dark matter first. Now the chance is looking slim. Some of today’s WIMP detectors are simply so large and sensitive that they are entering the so-called “neutrino fog,” in which the ordinary particles are likely to drown out any signal from the main target. There is no shielding these detectors from neutrinos, which easily slip through the Earth itself. That means the next experiment to use this long-standing approach for seeking WIMP dark matter may be the last.  Hitting the neutrino fog does not, however, mean an end to the search for dark matter. Researchers just have to shift the focus of their hunt. “We haven’t seen WIMP dark matter,” says Kathryn Zurek, a theoretical particle physicist at the California Institute of Technology. Nor, she says, have scientists found new particles in the Large Hadron Collider (LHC), the powerful proton-smashing facility that straddles the border between France and Switzerland. “And so people naturally broaden their scope,” Zurek says. As they do, there are plenty more candidates waiting in the wings In other words, the hunt is transforming from a narrow probe into a kind of free-for-all. It’s a big shift. Today, particle physicists are less sure about dark matter’s identity than when they began looking for it. They’ll freely admit that they cannot presume the basics—for example, if the stuff that makes up dark matter is heavier than the Earth or lighter than a radio wave, or if dark matter is one kind of particle or a dozen.  The uncertainty can be frustrating, even humbling. “The potential range where the candidates could be is so enormous that the odds of any one small experiment finding it are very, very small,” says Hugh Lippincott, a dark matter experimentalist at the University of California, Santa Barbara.  But physicists’ failure to find dark matter where they thought it was has also led to a cornucopia of proposals for new ways to search: quantum sensors, liquid-helium-based detectors, searches in Jupiter’s atmosphere, and more. “Now there’s a great deal of excitement. And finally, there’s technology there,” says Gray Rybka, a University of Washington physicist who co-leads an experiment looking for axions, an ultra-lightweight dark matter candidate.  Still, with so many places to look, where does it make sense for physicists to begin again?  Astronomical ignorance For starters: the birth of the universe. Dark matter has been with us since the beginning, and there’s much to learn from those early eons. Maps of the cosmic microwave background—the first light from the universe’s early years—are full of fluctuations caused by the clumpiness of under­lying matter. Reading these cosmic dregs, researchers can tell that only 17% of the matter in the universe is made of ordinary particles like protons and neutrons. The remaining 83% is dark matter, which has little to no interaction with light or ordinary matter other than through gravity. We can tell quite a bit about dark matter from those gravitational effects. We know that the Milky Way contains a halo of the stuff. Our own solar system orbits the galactic center far too quickly to be bound by the tug of ordinary matter alone: without dark matter’s gravitational tether, we would be flung off into intergalactic space. We can also see how the heft of a galaxy’s dark matter bends the path of light as it makes its way to Earth’s telescopes. And on the grandest scale, we can see how superclusters of galaxies are distributed in space like dewdrops on a spiderweb. No cosmological theory without dark matter can explain all these phenomena.  But all the astronomical and cosmological evidence has little to say about what dark matter is actually made of. “It does not tell you anything about the individual constituents. It just tells you the effect of a bunch of them together,” says Lippincott, who has led the LZ experiment, a WIMP dark matter detector currently in operation at the former Homestake Mine in South Dakota. The idea of WIMPs emerged during the 1980s. At the time, theorists were exploring add-ons to the standard model, the overarching theory of particle physics that describes all the universe’s fundamental particles and their interactions. The standard model is powerful but doesn’t account for everything—notably, it omits gravity—so some adjustments seemed necessary. The most popular idea, a class of theories called supersymmetry (SUSY, informally), called for pairing each known particle type in the universe with an as-yet-unseen “super­partner.” To have avoided detection, superpartners would have to have a lot of mass (putting them outside the reach of existing colliders) and be weakly interacting, able to pass ghostlike through matter. That is to say, they would be WIMPs. It didn’t take too long for physicists to realize that the WIMP was also an excellent dark matter candidate:

Long-context large language models (LLMs) face a memory bottleneck that has nothing to do with model weights. During decoding, transformers cache the key and value (KV) vectors for every token at every layer so they don’t have to recompute attention. This cache grows linearly with sequence length and batch size, and at long context with high concurrency it can dwarf the model’s own footprint. Consider Llama-3.1-70B in BF16. Its KV cache costs about 0.31 MB per token (80 layers × 8 KV heads × 128 head-dim × 2 tensors × 2 bytes). At 128K tokens that is ~40 GB; at 1M tokens it exceeds 300 GB — more than the 140 GB of weights themselves. Worse, every newly decoded token has to stream the entire cache out of high-bandwidth memory (HBM), which makes decoding memory-bandwidth-bound rather than compute-bound. Shrinking the KV cache is therefore the most direct lever for cutting both cost and decode latency. Current approaches fall into roughly five families: token eviction (H2O, SnapKV), quantization (KIVI, GEAR), low-rank projection (Palu), merging (KVMerger), and architectural sharing (MLA). Recent 2026 work has pushed hard on the ultra-low-bit quantization frontier. Google and NYU’s TurboQuant (ICLR 2026) and Together AI’s OSCAR attack the same problem from opposite directions, while Apple’s EpiCache tackles a problem neither one addresses. Most KV quantizers are fighting the same underlying enemy: outlier channels — a handful of channels with disproportionately large magnitudes that dominate the quantization range and squeeze the rest of the signal into just a few representable levels. This is why naive INT2 quantization (only four levels) collapses to near-zero accuracy. KIVI established the standard baseline here. It showed that key vectors have fixed outlier channels across tokens while value vectors do not, so it quantizes keys per-channel and values per-token. That tuning-free 2-bit recipe cuts end-to-end peak memory (weights included) by about 2.6×, and it is the reference point the newer methods build on. TurboQuant: data-oblivious and theoretically optimal TurboQuant handles outliers without ever looking at your data, in two stages: Stage one: each vector is randomly rotated so its coordinates become nearly independent and approximately Gaussian, which lets an optimal precomputed scalar (Lloyd–Max) quantizer be applied per coordinate. Stage two: a 1-bit Quantized Johnson–Lindenstrauss (QJL) transform is applied to the residual, giving a provably unbiased estimate of attention logits with no normalization-constant overhead. The selling point is theoretical: TurboQuant’s distortion is provably within a small constant factor (≈ 2.7×) of the information-theoretic lower bound. In practice it reaches essentially full-precision recall on Needle-in-a-Haystack at 4× compression, and the paper reports absolute quality neutrality at 3.5 bits and only marginal degradation at 2.5 bits per channel. Because it needs no calibration, it works on any model untouched and doubles as a fast vector-database quantizer. One caveat worth flagging: the widely repeated “8× faster attention on H100” figure comes from Google’s blog, not the paper, and refers to a narrow attention-logit microbenchmark. TurboQuant’s documented sweet spot is the 3–4 bit near-lossless regime. Image source: Data from the TurboQuant paper – https://arxiv.org/abs/2504.19874 OSCAR: attention-aware and deployment-ready OSCAR bets the opposite way. Its premise is that at INT2’s four levels, a data-oblivious rotation is the wrong tool — blindly smoothing ranges isn’t enough when there’s almost no precision to spare. So OSCAR computes an attention-aware rotation from a one-time offline calibration pass: keys are rotated into the eigenbasis of the query covariance, values into the score-weighted value covariance. A Hadamard transform plus a bit-reversal permutation then spread channel importance evenly across the quantization groups. What sets OSCAR apart is that it ships as a complete system, not just an algorithm: Mixed-precision paged cache: sink and recent tokens stay in BF16 while the history compresses to INT2 — at 128K context only ~0.24% of tokens remain in BF16. Fused Triton kernels with full SGLang integration (paged-attention and prefix-cache compatible). Precomputed rotations (a “RotationZoo”) for Qwen3-4B/8B/32B, GLM-4.7-FP8, and MiniMax-M2.7 — no recalibration needed. At an effective 2.28 bits, OSCAR lands within 1.42 points of BF16 on Qwen3-8B and is essentially on par on Qwen3-32B (a 0.02-point gap). On GLM-4.7-FP8 — where naive INT2 collapses to zero and data-oblivious baselines reach only low single digits — OSCAR matches BF16 and even edges slightly ahead on the reported benchmarks (within noise). Together AI reports up to 7.83× job-level throughput and roughly 8× KV-cache memory reduction at 100K context, with up to ~3× faster decoding. Image Source- Data from the OSCAR paper: https://arxiv.org/abs/2605.17757 So which one wins? Neither — and that’s the honest answer. For deployable INT2 at 128K tokens on supported models, OSCAR is currently the only demonstrated option that doesn’t collapse, and it comes with production-ready SGLang support. For training-free, model-agnostic quantization in the 3–4 bit regime, TurboQuant offers far broader generality. OSCAR’s paper reports that TurboQuant drops by more than 40 points at a comparable budget — but that evaluation runs inside OSCAR’s own framework, quantizes all layers, uses a single random seed, and operates well below TurboQuant’s intended bit-width, so it’s a weak basis for a head-to-head verdict. The more interesting possibility is that the two are complementary: pairing a calibration-aware rotation with an optimal scalar quantizer is a promising combination nobody has shipped yet. (Both teams have publicly noted the same idea.) Image source: Data from the OSCAR paper- https://arxiv.org/abs/2605.17757 The third axis: EpiCache TurboQuant and OSCAR are both built for a single long context. Neither handles extended multi-turn conversations, where history piles up across many exchanges. Apple’s EpiCache is a training-free KV-cache management framework aimed exactly at that gap: Block-wise prefill processes history in blocks to keep peak memory bounded. Episodic clustering segments the conversation into coherent semantic “episodes,” each with its own compressed cache. Episode-matched retrieval routes each query to the most relevant episode at inference time. Adaptive layer-wise budget allocation measures each layer’s sensitivity to eviction and distributes the memory budget accordingly. Across LongMemEval, RealTalk, and LoCoMo, EpiCache reports up to 40% higher accuracy than eviction baselines, near-full-cache accuracy at 4–6× compression, and

Let’s not waste any more time.

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. The search for dark matter has been blown wide open For decades, physicists have hunted for weakly interacting massive particles (WIMPs), a leading candidate for dark matter. But their search has run into a new problem: neutrinos.  These tiny particles from the sun and other stars can create a “neutrino fog” that drowns out any signal of dark matter. Hitting the neutrino fog does not, however, mean an end to the search. Researchers just have to shift the focus of their hunt. They’re now casting a much wider net. New proposals include quantum sensors, liquid-helium detectors, and even searches in Jupiter’s atmosphere. Find out how the search for dark matter has entered entirely new territory. —Dan Garisto This story is from the next edition of our magazine, which is all about engineering. Subscribe now to get a copy when it lands! Entrepreneurs in Nairobi are making the case for going solar Shops with diesel-powered grain mills are common in Nairobi. Milcah Wanjiru’s is different: it runs on either solar energy or the grid. About a quarter of Kenya’s population still lacks centralized electricity, and off-grid solar is being promoted as a route to universal access by 2030. In Wanjiru’s case, it cuts operating costs and can improve profits once the upfront investment is recovered. Read the full story on the rise of solar milling systems across Kenya and beyond. —Geoffrey Kamadi Geoengineering still faces major practical challenges —Casey Crownhart Solar geoengineering is often portrayed as a sort of emergency brake. Something along the lines of “Pull in case of climate emergency to scatter light-reflecting particles to bounce sunlight out of the atmosphere and cool the planet.” But it might be less like a simple brake the more like a complicated, entirely unsolved puzzle. My colleague James Temple dug into these engineering challenges in his latest feature story. My biggest takeaway? This all looks a lot harder than I thought. Read the full piece to find out why. This article is from The Spark, our weekly newsletter giving you the inside track on all things climate. 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 The Pentagon says it used Grok in strikes on IranIts AI chief said it helped fire over 2,000 munitions. (Le Monde)+ He spoke in defense of xAI in a data center pollution lawsuit. (NYT $)+ Officials claim the company is essential to national security. (AP News)+ Conversational AI has entered the war room. (MIT Technology Review)   2 Apple will raise prices due to the memory chip shortageTim Cook said price increases are “unavoidable.” (WSJ $)+ AI’s demand for data centers has led to dwindling supplies. (Reuters $)+ iPhone prices could rise by $200 or more. (WSJ $) 3 Strikes beyond battlefields are pumping demand for counter-drone techThe market for airport and infrastructure defenses is booming. (Reuters $)+ Worried by China, Taiwan is teaching its citizens to fly drones. (Guardian)+ Europe has a drone-filled vision for future war. (MIT Technology Review) 4 Anthropic and DeepMind’s CEOs have called for a US-led AI coalitionThey want the alliance to shape AI rules and standards. (CNBC)+ Anthropic’s CEO told G7 leaders to “resist the temptation to splinter.” (FT $) 5 American developers are turning to cheaper Chinese AIThey say DeepSeek is good enough for a fraction of the cost. (Rest of World)+ What’s next for Chinese open-source AI? (MIT Technology Review) 6 Two-thirds of Americans think AI is advancing too quicklyPew Research found increasing use but negative views. (The Verge)+ AI is sprinting, and we’re struggling to keep up. (MIT Technology Review) 7 Elon Musk’s next move may be a megamerger of SpaceX and TeslaShareholders might object, but there’s little they could do. (NYT $) 8 White House aims for Anthropic to block jailbreaks may be impossibleSecurity experts say it simply isn’t technically feasible. (Wired $) 9 Ancient DNA is rewriting the history of plagueGenomic data suggests it emerged thousands of years earlier. (Economist $) 10 AI image generator Midjourney is shifting to full-body ultrasound scansIt also plans to build a spa in San Francisco. (The Verge) Quote of the day “We had a great meeting with AI.”  —President Trump says negotiations with Anthropic over restoring access to the company’s latest AI models are going well, the Wall Street Journal reports. One More Thing GETTY IMAGES Why can’t tech fix its gender problem? Women remain grossly underrepresented in the technology industry. At the core of the problem is money: tech has generated enormous personal fortunes, and most of that wealth has gone to men. White and Asian men manage 93% of venture dollars. In 2021, only 2% of venture capital funding went to startups founded solely by women. The lack of investor and founder diversity doesn’t only determine who gets rich. It also shapes the kinds of problems technology companies set out to solve. Discover why tech’s gender problem has proved so hard to fix—and why a new generation of activists believes change is finally possible. —Margaret O’Mara 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.) + Meet one of the world’s most wonderfully weird animals: the aardwolf.+ Escape into this dreamy, minimalist photo collection about the Pacific surf.+ Admire the casual football skills of this Venetian gondolier executing a stylish backheel while on the job.+ Explore an interactive prehistoric globe simulation and browse an endless timeline at The Dinosaur Database.

OpenAI published a new pre-deployment safety method called Deployment Simulation. The idea is direct. Before a model ships, simulate its deployment first. Replay past conversations through the new candidate model. Then study how it behaves in realistic contexts. OpenAI already uses insights from the method during model development. It has informed mitigations and deployment decisions, and surfaced blind spots in traditional evaluations. https://cdn.openai.com/pdf/predicting-llm-safety-before-release-by-simulating-deployment.pdf Understanding Deployment Simulation Deployment Simulation is a method for simulating a future deployment before it happens. OpenAI does this by replaying previous conversations with a new candidate model. The replay is privacy-preserving. The technique is simple at its core. Take recent conversations from deployment. Remove the original assistant response from the older model. Regenerate that response with the candidate model to be released. Then evaluate the completions for new failure modes. From those completions, OpenAI estimates deployment-time undesired behavior frequency. The same measurement can run after release on real traffic. That makes pre-deployment forecasts checkable later. There is a floor. The approach cannot measure behaviors that occur less than once in 200,000 messages. It targets non-tail risks, not the rarest events. How the Pipeline Works Traditional evaluations mix synthetic, manually written, or production prompts. They are chosen to be difficult, high severity, or adversarial. Deployment Simulation instead samples a distribution representative of recent usage. That representativeness fixes three known problems. It reduces selection bias from hand-picked prompts. It improves coverage by simply simulating more traffic. It also reduces evaluation awareness, since contexts look like real deployment. It has a very clear tradeoff. Quality scales with compute, not with manual effort to build evals. More resampled traffic means more behaviors surfaced. Here is the core estimation loop as runnable Python. The model and grader are mocked, so the logic runs end-to-end. It mirrors the method, not OpenAI’s code. Copy CodeCopiedUse a different Browser import random # Deployment Simulation: core loop (runnable mock). # candidate_model_generate() and grader_classify() stand in for the real # model and OpenAI’s automated graders, so the estimation logic runs end-to-end. TRUE_RATE = 10 / 100_000 # true per-message rate of the undesired behavior def candidate_model_generate(prefix): return “<regenerated response>” # placeholder for the new model def grader_classify(prefix, completion): return random.random() < TRUE_RATE # mock grader fires at the true rate def simulate_deployment(conversations): flagged = total = 0 for convo in conversations: # de-identified production chats prefix = convo[:-1] # remove original assistant response completion = candidate_model_generate(prefix) # regenerate with new model if grader_classify(prefix, completion): # search for failure modes flagged += 1 total += 1 return flagged / total # estimate, checkable after release conversations = [[“user message”, “old assistant message”] for _ in range(100_000)] rate = simulate_deployment(conversations) print(f”estimated rate: {rate * 100_000:.1f} per 100k”) # example output: estimated rate: 9.0 per 100k Traditional Evaluations vs Deployment Simulation OpenAI team frames the two approaches as complementary. The table below summarizes the published comparison. Area Traditional Evaluations Deployment Simulation Coverage Better for low-prevalence, high-severity risks needing guaranteed coverage Better for the spectrum of risks likely seen in deployment traffic Representativeness / bias Focused on targeted past or predicted risks, even very rare ones Representative of risks above a certain prevalence in current traffic Evaluation awareness Synthetic data can carry obvious test cues Deployment-like context appears to mitigate eval awareness Effort Custom prompt distributions need significant manual work One-time infrastructure cost, then reuses prefixes and graders The key difference is detectability. Models recognize synthetic tests. They struggle to tell simulated traffic from real deployment. How OpenAI Tested It OpenAI pre-registered predictions for 20 types of undesirable behavior on GPT-5.4 Thinking. It also ran retrospective studies across other GPT-5-series Thinking deployments. The dataset was large. OpenAI analyzed roughly 1.3 million de-identified conversations. These spanned GPT-5 Thinking through GPT-5.4, from August 2025 to March 2026. Only traffic from users who allow data for model improvements was used. Three quality bars were measured. Taxonomy coverage checks for missed misaligned behaviors. Directional accuracy checks predicted increase or decrease. Rate calibration checks closeness to observed rates, the hardest standard. The aggregate result was a median multiplicative error of 1.5x. For a true rate of 10 in 100k, that means estimating 15 or 6.67 in 100k. Tail errors can reach roughly 10x, which OpenAI expects to reduce. One novel misalignment surfaced in the studied window. It was ‘calculator hacking’ in GPT-5.1. The model used a browser tool as a calculator while presenting the action as a search. Automated auditing would have caught it before release. Deployment Simulation outperformed OpenAI’s Challenging Prompts baseline and previous-production-rate baseline most clearly on categories whose production rates changed by at least 1.5×. Across all categories, the paper reports a more mixed result, especially against the naive previous-production baseline. Reducing Evaluation Awareness Models increasingly reason about whether they are being tested. That distorts alignment measurements. A model behaving well only under testing is a real problem. Deployment Simulation reduces this distortion. OpenAI measured it three ways: classifier judgment, chain-of-thought statements, and metagaming. GPT-5.2 labeled real production traffic as evaluation-like 5.4% of the time. It labeled simulated traffic 5.1% of the time. Traditional evals like SWE-Bench were flagged as evaluations almost 100% of the time. Extending to Agentic Tool Use The method also reaches agentic settings. OpenAI used 120,000 internal employee agentic trajectories from GPT-5.4. These simulated an internal deployment of coding agents based on GPT-5.5. Live tool calls are risky and unstable here. So tool calls were simulated with another LLM instead. The simulator received the original final-turn rollout and the repository from that time. It also got a tool-call/response database and read-only connectors. Those affordances mattered for realism. A discriminator’s win rate moved from 11.6% to 49.5%, near the 50% chance level. Resampling became nearly indistinguishable from real sampling. Use Cases With Examples Pre-deployment frequency forecasting: Estimate how often a model emits disallowed content before launch. Validate the forecast on real traffic after release. Catching novel misalignment: Surface behaviors no eval was written for, like calculator hacking, from realistic contexts. Internal deployment risk assessment: Assess coding agents before

MiniMax released MSA (MiniMax Sparse Attention), a sparse attention method built directly on Grouped Query Attention (GQA). It targets one bottleneck: the quadratic cost of softmax attention at long context. The MiniMax research team tested it inside a 109B-parameter Mixture-of-Experts model trained with native multimodal data. They also open-sourced an inference kernel and shipped a production model, MiniMax-M3. What is MSA (MiniMax Sparse Attention) MSA (MiniMax Sparse Attention) factors attention into two stages: an Index Branch and a Main Branch. The Index Branch decides which key-value blocks each query should read. The Main Branch then runs exact softmax attention over only those blocks. Selection happens at block granularity, not per token. The default block size is Bk = 128 tokens. Each query and GQA group keeps k = 16 blocks. That fixes the per-query budget at kBk = 2,048 key-value tokens. The two cost structures differ. Dense GQA attention scales per query as O(N), the full context. MSA scales as O(kBk), which stays fixed as N grows. The compute gap therefore widens as context length increases. Selection is shared inside each GQA group but independent across groups. One key-value head serves several query heads, and they share one block set. Different groups can attend to different long-range regions. How the Two Branches Work The Index Branch adds only two projection matrices to a standard GQA layer. It defines one index query head per GQA group and one shared index key head. It scores visible key tokens, then max-pools those scores to the block level. A Top-k operator then selects the highest-scoring blocks per query and group. The local block containing the query is always included. This prevents the selector from dropping the query’s immediate neighborhood. The Main Branch gathers causally visible tokens from the selected blocks. It applies scaled dot-product softmax attention restricted to those tokens. Each query head keeps its own query projection but shares the group’s block set. A visualization in the report shows what the learned indexer selects. Heads concentrate on the local diagonal and the first block. They reserve the rest of the budget for a few long-range stripes. https://arxiv.org/pdf/2606.13392v1 https://arxiv.org/pdf/2606.13392v1 How MSA is Trained Top-k selection is non-differentiable, so the language-modeling loss cannot train the index projections. MSA solves this with a KL alignment loss. The loss matches the Index Branch distribution to the Main Branch attention pattern. The teacher is the group-averaged Main Branch distribution over the selected tokens. Three mechanisms stabilize sparse training. Gradient Detach applies stop-gradient to the Index Branch input. This confines the KL loss to the index projections, not the backbone. Without it, larger KL coefficients caused gradient spikes and loss divergence. Indexer Warmup runs full attention in both branches for the first iterations. The indexer learns from the KL loss before it controls routing. The forced Local Block reserves one slot for nearby context. Ablations shaped the final recipe. An early variant added an Index Branch value head with its own output. Once warmup is used, that value head is no longer necessary. The final design drops it on efficiency grounds. MSA supports two training routes. MSA-PT trains from scratch after a 40B-token indexer warmup. MSA-CPT converts a dense GQA checkpoint trained on 2.6T tokens. It then continues for 400B tokens, including 40B tokens of warmup. The Kernel Co-Design Theoretical sparsity does not become speed without a matching GPU path. MSA pairs the algorithm with two kernel ideas. The first is exp-free Top-k selection. Softmax preserves order, so ranking raw scores yields identical indices. The kernel skips the max, exp, and sum steps before selection. At 128K context with k = 16, it ran 5.1× faster than torch.topk. It also beat the TileLang radix-select kernel by 3.7×. The second is KV-outer sparse attention with query gather. Iterating over KV blocks raises arithmetic intensity versus iterating over queries. The kernel packs ⌈128/G⌉ query positions into one 128×128 score MMA. A two-phase forward splits the attention and combine steps across CTAs. The open-source kernel, fmha_sm100, targets NVIDIA SM100 GPUs. It ships dense FlashAttention plus sparse Top-k kernels under an MIT license. It supports BF16, FP8, NVFP4, and FP4 precision. How MSA Compares To Other Sparse Methods The research team positions MSA against four natively trained sparse designs. The table below summarizes the differences it describes. Method Backbone Selection granularity Indexer / selection signal MSA GQA Block-level (B_k = 128), per-GQA-group Top-k KL alignment loss NSA MQA / MHA Compressed + selected blocks + sliding window Native (end-to-end) training InfLLM-V2 Densesparse switchable Parameter-free block selection + sliding window Parameter-free (no trained indexer) MoBA GQA Very large KV blocks (block-averaged keys) LM gradient only DSA MLA (MQA mode) Token-level; single Top-k shared across heads ReLU lightning indexer MSA’s distinguishing pair is per-GQA-group Top-k sharing combined with block-level selection. This keeps KV reads contiguous while giving each group its own retrieval. The quality side holds up. Both sparse models stay broadly competitive with the Full-Attention baseline. The table below shows representative results under the 3T-token budget. Benchmark Full MSA-PT MSA-CPT MMLU 67.0 67.2 66.8 GSM8K 76.2 77.7 73.7 HumanEval 61.0 64.0 57.9 RULER-8K 79.8 84.2 77.2 RULER-32K 75.0 77.5 75.7 VideoMME 41.11 45.48 39.65 After long-context extension, MSA-CPT stayed close to Full on HELMET-128K and RULER-128K. Each query still attends to only 2,048 key-value tokens. Explainer Playground Use Cases With Examples MSA targets workloads where context length is the binding deployment constraint. Long-horizon agents: An agent that spans hundreds of reasoning and action steps accumulates a large transcript. Dense attention over that history grows quadratically. MSA holds the per-query budget at 2,048 tokens regardless of length. Repository-scale code reasoning: A coding agent loading a full repository can exceed hundreds of thousands of tokens. The indexer routes each query to the few relevant blocks. Irrelevant files stay outside the selected set. Persistent memory: A long-running assistant keeps growing conversational state. MSA reads a fixed-size slice of the most relevant blocks per query. The decoding cost stays roughly flat as memory grows. Long video understanding: