well well well rumor from an NYU math professor that openai's huge math drop yesterday was only batch 1 of 3 which means there are still 2 more batches left, and the first batch alone included 722 manuscripts across 372 result families from their internal frontier model

Scientists are losing their minds over this paper. They proved every mitochondria in your body is a tiny quantum device running at 71 THz. And it responds to light. Researchers detected a mysterious 71.0-terahertz vibration signal. this frequency is only present in healthy, living cells and tissues. exact moment the cell dies or the mitochondrial structures are broken apart, the signal vanishes completely. How does this happen? Classical physics has always struggled to explain the high energy-efficiency of life. So the team built a quantum model of light-matter coupling and discovered something called a "mito-polariton". It is a literal quantum superposition state that forms when light couples with lipid ch2 bonds inside functional mitochondria. This quantum resonance splits the intrinsic vibration mode into two levels, creating that bizarre 71 thz frequency that you only see in living things. Why does this matter? Because this quantum state actually acts as a highly efficient channel to modulate atp production in living cells. our bodies literally use quantum mechanics to generate the energy that keeps us alive.

“okay, so what? it’s just math.” no, you don't get it. here’s what OpenAI’s math breakthroughs could eventually help make possible in the real world: Vlasov–Maxwell → stronger foundations for fusion research, whose ultimate prize is virtually limitless clean energy to power civilization Calderón’s Problem → portable body scanners using electrical signals, making medical imaging cheaper and easier to access Maximum-Cardinality Matching → faster searches for compatible kidney swaps across large donor pools, helping hospitals coordinate lifesaving transplants Inverse Elasticity Problem → scans that map tissue stiffness, helping doctors locate suspicious growths inside the body Mumford–Shah Conjecture → better tools for spotting tissue changes in brain scans, helping doctors identify signs of disease Bose–Einstein Condensation → stronger foundations for quantum sensors that could help vehicles navigate without GPS Simple Stochastic Games → better safety checks for self-driving cars and robots before dangerous mistakes reach the real world Matrix Multiplication → cheaper AI and bigger scientific simulations using the same computers Edit Distance → faster DNA comparisons, helping researchers study genetic changes linked to disease The k-Server Problem → robots that waste less movement and energy, making warehouses more efficient and goods cheaper to move a reminder that math is the foundation of science. so accelerating mathematical discovery could compress centuries of scientific progress into years. insane timeline to be alive for!

BREAKING: SpaceX wins FCC approval for its 15,000-satellite Direct-to-Cell constellation. The order also includes a conditional U.S. waiver of satellite power limits, advancing SpaceX’s push to expand mobile coverage beyond the reach of cell towers.

The DEA has released its updated report confirming what advocates and researchers have pointed out for years: there has never been a reported cannabis overdose death. Meanwhile, the CDC reports about 178,000 people die from excessive drinking each year. marijuanamoment.net/there-ar…

Ok so I took a closer look at the results, and OpenAIs AI-generated mathematics manuscripts are *even more* significant than I initially thought. I spent the morning going through it. Some thoughts. The list is absurd. A zero-free half-plane for the zeta function (Re s > 7/8), which is the first result of its kind in over a century. Hilbert's tenth problem over the rationals. The Hodge conjecture for CM abelian varieties. Irrationality of Catalan's constant. Dozens more. Any one of these would normally be a career. But the number that many arent seeing is the following: It's 3. That's the average hours of ChatGPT Pro compute per result. A month ago, Navier–Stokes took them around 10,000 agents and 88 hours. That efficency gain within just a few weeks. Also OpenAI claims to have solved the quasi-Riemann hypothesis. That alone would be a historic breakthrough in mathematics. This is a weaker version of the famous Riemann hypothesis, which concerns how prime numbers are distributed. The full hypothesis remains unsolved, but the claimed advance would be enormous in its own right. Math twitter obviously is shocked. Again: this is literally the intelligence explosion happening right now. 2027 will be the year of Superintelligence. Im now convinced by that.

Important note regarding Grok @Bot: Going forward, @SpaceX will use the best back end model for any given task, including Claude Opus 5.5, MidJourney, Suno and other leading APIs. Whatever is most likely to give you the best outcome.

Insane developments have been happening on Qwen 3.8 Flash. In 2 weeks, suddenly everyones computer can run Qwen 3.8 Flash. I wanted to find out what happened and why everyone in local AI is using this model now. Every number below is a receipt I verified, my own hardware or community, and the ones from my fleet live in my bench repo. Here is what I found. WHY THIS MODEL Qwen3.8-Flash-Next is a 180B parameter MoE that only activates around 6B per token. On Artificial Analysis it scores 40 on the intelligence index against Claude Opus 4.8 at 42. It beats Opus on Terminal-Bench 4.0 (25 percent against 22), AutomationBench (56 against 46), and GDPval. Opus still wins Humanity's Last Exam (49 against 38) and SWE-bench Pro (69.2 against 62.5, each vendor's own harness). My read: Opus-class on agentic and coding work, one step behind on the hardest reasoning. People call it Qwen 3.8 Flash. Same model. WHY IT RUNS ON ALMOST ANYTHING EXPLAINED Traditional LLM understanding is everything has to run on vram alone. Model doesn't fit VRAM - you can't run it. The Chinese AI companies broke the conventional datacenter model basically, by showing people off-loading critical parts of the model out of it is possible, but there was a catch - Everyone knows that RAM offload is slow, and it was true. A dense 70B at 4-bit is around 35 GB of weights and every token touches all of it. From system RAM at 60 GB/s that is about 1.7 tok/s. Flash-Next breaks the rule because of its shape. 512 experts per layer. The router picks 10, plus one always-on shared expert. Per token you read maybe 2 percent of the expert bank, around 2.2 GB at 3-bit. Same RAM, same bandwidth: 60 divided by 2.2 is around 27 tok/s as a pure-RAM streaming bound. Result: Laptop runs suddenly become possible. Sparse active set + Treat your whole computer as RAM. The new engines stack three things on top: The cache hierarchy. Expert usage is lopsided. Some fire constantly, most rarely. The engine treats your PC like a CPU cache: hot experts pinned in VRAM, warm ones in RAM where the CPU computes them, cold ones streamed from SSD. Speed becomes hit rate, not bandwidth. The n-gram table. 51B of the 180B parameters are a lookup table. Not matrix math, an index. It sits on your SSD and streams a few rows at a time. MTP speculative decoding. A draft head guesses, the big model verifies several tokens per pass, and corrrect guess is a free speed up. STRATA CHANGED THE GAME (AGAIN) The engine that went viral is Strata. MIT-licensed, built for this one model family, one-click install. It serves an Anthropic-compatible API on localhost, so Claude Code and Codex CLI point straight at it. That is the real adoption driver. Same 5090, llama.cpp: 15 tok/s. Strata: 160 tok/s. Engine, not model. The quant side moved too. ISTA-DASLab's GSQ-RCO assigns bits per tensor by measured sensitivity. Their 3.5-bit build (83.6 GB) scores near or above the BF16 base on most evals. A quarter of the size, matching quality. THE FULL MAP 8 GB VRAM RTX 4060 laptop (Strata, Q3): 30 tok/s at 192k context, 130W from the wall. Tono_Ken3's receipt. A gaming laptop is now an inference server. GTX 1080: Apparently this runs at 20 tok/s painfully, but it works lol 12 GB VRAM RTX 5070 + 64 GB RAM (Strata, Q2_0): 94 tok/s decode, 2,650 tok/s prefill at 32k. The project's reference rig. 24 GB VRAM RTX 3090 + 64 GB RAM: Two lanes. llama.cpp UD-Q2_K_XL with expert offload: 28 tok/s at 65k context, sustained over 1.59M production tokens (eirrann_art). EXL3 2.5bpw with MTP: 38.6 tok/s, full 262k native context, needs around 59 GB host RAM (r0b0tlab). RTX 4090 + 64 GB RAM (Strata, IQ3_S): 60 tok/s decode, 2,500 tok/s prefill. 5000e12's receipt. RTX 5090 (Strata, IQ3_S): 160 tok/s, up from 15 on llama.cpp. EpicMaan's receipt. 32 GB VRAM (LEGACY) 2x V100 32 GB (64 GB total): 60 to 88 tok/s at 262k context on pentacoxian-dev's custom quant. ryu15's receipt. The e-waste lane is real. 128 GB UNIFIED (DGX SPARK) 1x Spark (my grid, EXL3 native): 107 tok/s count, 71.7 prose, 15 of 15 needle at 243k context. 2x Spark TP2 (my grid): 68 tok/s count on stock NVFP4. 138 on hibrid48's stock-incompatible fork. Same chips, the recipe alone is worth 2x. 64 GB+ UNIFIED (MAC) Macs: two engines carry it, not just the GGUF lane. TensorFold runs the full 180B on M1 through M4 using oQ checkpoints, with the n-gram table scaled and dense projections moved to the matrix units before M5. oMLX, which owns the oQ format, serves it too, with SSD-paged KV caching and a menu-bar app. Floor for the 180B is still 64 GB unified. Under that, the Mac lane is the Qwen3.8-27B dense: 131-160 tok/s on an M5 Ultra, 16 streams at 32k on a 64 GB machine. THE 32 GB QUESTION Can it run on 32 GB of RAM? Two weeks ago the answer was no, the floor was 48 GB. Strata's spec floor is now 32 GB RAM plus 12 GB VRAM, but every receipt I could verify sits at 64 GB. The floor claim moved to 32 this month. The receipts live at 64. If you are on 32, the honest answer today is the Qwen3.8-27B dense. This model will become a mainstay for local AI for many months to come. It's reasoning capabilities are very good in my own testing, it's not GLM 5.3 or Deepseek V4.1 level, but for a driver for most daily use. Opus class model intelligence has arrived to almost every computer, and yes Qwen 3.8 Flash is stronger than Qwen 3.8 27B dense in my testing. Sources, receipts, and how tos in the replies 👇

We’re releasing a broad range of new mathematical results produced by an internal frontier model. We’ve been consulting with the independent Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study, and we have drawn on their advice and public recommendations to inform how we release these results. github.com/openai/math

This guy re-created all the adobe apps with opus-5.5, ported them to rust, and opensourced them something really cool is happening

holy fucking shit. 90+ opus 5.5 agents spent three days doing actual materials research and surfaced two candidates for a class of room-temperature magnetic semiconductor researchers have been searching for. and one of them already fucking existed. humans made it. studied it. published papers on it. it was literally synthesized in 1999. then 27 years later a swarm of opus 5.5 agents comes along and goes “wait, this might have the exact properties everyone has been looking for.” there must be an unbelievable amount of science hiding in plain sight simply because no human has connected the right papers, calculations and ideas yet.

HOLY, the rumors were true: OpenAI has published 722 mathematical manuscripts produced by an *unreleased* internal model. The collection groups them into 372 families of related results, drawn from an evaluation involving approximately 4,000 research problems. OpenAI says the standard procedure used an average of three hours of ChatGPT Pro thinking compute per result. The release includes papers, proof artifacts and selected reasoning summaries. The model itself remains unreleased.

THIS IS BIG! The Cancer Cells That Changed Their Minds A KAIST team just reversed cancer without killing a single cell — and the method may matter more than the result. For a century, oncology has had exactly one strategy: find the cancer and destroy it. Chemotherapy poisons it. Radiation burns it. Surgery cuts it out. Immunotherapy teaches the body to hunt it. Every weapon differs in precision, but not in philosophy. The tumor is an enemy. You win by killing. Now a team at South Korea's KAIST has proposed a heresy: *what if the tumor doesn't need to die?* In a study led by Professor Kwang-Hyun Cho of the Department of Bio and Brain Engineering, published in Advanced Science ("Control of Cellular Differentiation Trajectories for Cancer Reversion," DOI 10.1002/advs.202402132), researchers took colon cancer cells and — without poisoning, irradiating, or cutting them — turned them back into normal cells. Not dead. Not damaged. Just normal again. The tumors, grown in mice, shrank dramatically. The surrounding tissue was left intact, because there was never an attack to survive. This is early research — cell lines and animal models, no human trials, real obstacles still unsolved. But the conceptual break is the story. For the first time, cancer treatment has a second verb. Not just destroy. Also: convert. The digital twin of a cell The KAIST team's insight begins with a redefinition of what cancer is. The conventional view treats cancer as a pile of broken machinery: mutations accumulate, checkpoints fail, cells proliferate. Cho's group looked at the same evidence and saw something different — a trajectory. During oncogenesis, they observed, normal cells don't just break. They regress, sliding backward along the differentiation path they followed when they matured. A colon cell becomes, in effect, a confused stem cell: immature, proliferative, lost. If cancer is a wrong turn on a developmental road, then the treatment question changes. You don't blow up the road. You build a map and find the turn. That map is what the team calls a digital twin — a complete computational model of the gene network governing a cell's differentiation. Using data from 4,252 intestinal cells, they reconstructed a network of 522 interacting components, capturing how genes regulate one another as a cell matures or degrades. Then they did the audacious thing: they asked the simulation which levers, flipped together, would push a cancer cell back down the road toward normalcy. The answer came through a system they built called BENEIN (Boolean Network Inference and Control), which models gene interactions as logical relationships and systematically tests which interventions redirect the network's state. The simulation pointed to three master regulators — the genes MYB, HDAC2, and FOXA2 — acting together as the switch that holds the cancerous state in place. Turn all three off simultaneously, the model predicted, and the cell would stop proliferating and differentiate into something resembling a normal intestinal cell. They tested it, and it worked. In three colon cancer cell lines, suppressing MYB, HDAC2, and FOXA2 together strongly induced differentiation into normal-like cells. The cancer cells began expressing markers of healthy intestinal tissue. Proliferation collapsed — not because the cells died, but because they grew up. In animal models, tumors formed from the reprogrammed cells were dramatically smaller than controls, and under the microscope they looked far more like normal tissue. The signature achievement, and the one that would matter most to any patient reading this: there was no collateral damage. Nothing was poisoned, burned, or irradiated. The healthy tissue never came under fire, because there was no fire. 1 of 2

I'm open sourcing Leviathan: an indexer that lets AI agents search a database of any size without reading it. At 1M records it hands your agent 436 tokens instead of 107,000, and it finds the answer 99% of the time. github.com/elstongun/leviath…

Scorpion venom is valued at roughly $39 million per gallon (3.8 liters), making it the most expensive liquid on Earth. In 2-3 weeks, the scorpions will provide another drop.

Hark launches this week. The first 100,000 sign-ups get a paid plan for free I don't do anything without Hark anymore. The team has obsessed over every detail, and it shows. Can't wait to see how you all use it Join the waitlist: hark.com

"Most people aren't looking to save time, they're looking for ways to spend their time." 9 of 15 consumer internet categories have zero AI products in the Top 100. These built some of the biggest companies of the last two eras: - Streaming - Social - Dating - Gaming - Travel - Retail - Finance - Real estate - Jobs More charts in our Top 100 Consumer AI Apps breakdown: a16z.news/p/top-100-consumer…

Text watermarking in ChatGPT and Codex is rolling out. The new development is that, unlike Anthropic and Google, OpenAI (for now) is launching their watermarking - textGrain - 𝘰𝘯𝘭𝘺 in the EU, not globally.

Holy sh*t, Anthropic’s subscriptions offer so much more value, even compared to the new and efficient GPT-6.1 Sol. It’s not even close. SemiAnalysis puts Opus 5.5 at over 5x the API-equivalent value of GPT-6.1 Sol for agentic workloads at the same subscription price. It’s not even close on this measure. And on top: OpenAI just halved the usage limits on its $200 ChatGPT plan.

Another reason why local AI is absolutely necessary.