DeepSeek V4.1 Flash, Gemini 4 Carbon, and the exponential time hypothesis: 7 videos from October 26–November 1

The short version

OpenAI Claims It Just Did The Impossible

• The exponential time hypothesis is the assumption that 3-SAT, the standard hard problem, needs exponential time in the worst case. A generation of algorithms research has been built on that assumption, and a 3-SAT instance of modest size can take a quadrillion years to check by brute force. 1
• OpenAI's paper, dated October 17 and revised October 25, gives a deterministic algorithm that counts the satisfying assignments of any k-CNF formula in subexponential time — about 2^(n/(log n)²) bit operations — and polynomial space, and its abstract says plainly that the algorithm refutes the hypothesis. 2
• The paper hands over the algorithm itself, and it returns how many solutions exist rather than a yes or no. Zsolnai-Fehér spends a minute on why that matters for a video: counting is the harder task, and it is what makes sampling and approximation follow. 1
• The reach comes from Karp reductions. Routing medicine deliveries on a short fuel budget, assigning radio frequencies to thousands of transmitters, and choosing the fewest sensors for a site are all the same problem wearing different clothes, so speed on 3-SAT travels to each of them. 1
• Two limits are stated in the video. P versus NP is untouched, and the proof has not finished expert review; Zsolnai-Fehér calls it increasingly convincing while the formal review is still running. 1

Google "Carbon" just won everything...

• Carbon is Google's next frontier model, announced October 28 and ranked first on version 5 of the Artificial Analysis Intelligence Index at 59. Google's pitch is cybersecurity and long-horizon engineering rather than chat. 3
• The cyber numbers carry the claim: 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym for exploitation, and 71.0% on CWE-bench for patching, all reported as state of the art. 3
• You cannot use it. Google is releasing Carbon first to a subset of cyber defenders, then to trusted partners through its Fairwind program, while taking part in the US government's pre-release testing, before developers and consumers get access. Berman's objection is the gap between announcing and shipping. 3
• The price is the introductory $2 per million input tokens and $10 per million output, with cached input at 95% off, rising to $4 and $20 after the introductory period. Google does not say when that period ends. 3
• Berman's own comparison is about tokens per task rather than price per token: he reads Carbon at about 123,000 tokens for a benchmark task against 285,000 for Sonnet 5.5, and puts GPT-6.1 Soul Max at 68,000, the best intelligence density he can find. 4

Deepseek crushed everyone

• DeepSeek treats this as a small step and calls it a Pareto improvement in performance and cost. The notes credit reinforcement learning with self-feedback for the gains, and publishes weights on Hugging Face and ModelScope. 5
• The API carries a 1M-token context window, up to 384K output tokens and image input. List price runs $0.15 per million input and $0.60 per million output off-peak, doubling at peak, with cache hits at $0.003 and $0.006. 6
• Berman's demos are the reason to watch. A 3D mechanical watch page with hoverable parts, a Fall Guys clone, a foldable iPhone modelled from the web, and a count of 137 pastries on a counter that the model extends to a 130–145 range by reasoning about what is stacked behind. 7
• He also ran it locally on a DGX Station, where it built a working Excel clone in 50 minutes and a PDF reader called Folio. He reads its cost per completed task as 26 cents, against $2.11 for Kimi K3.1 Pro and $7.63 for Claude Fable 5.1 Max. 7
• The speed numbers come from third-party providers rather than DeepSeek: he cites 845 tokens per second from one and 1,286 from another. His conclusion is that quality, price and speed together make it the best overall model available today. 7

DeepSeek's New AI Is An Absolute Monster

• His answer is learning from self-feedback, which he offers as his own reading of DeepSeek's wording. The release notes themselves only name reinforcement learning with self-feedback as the method used for building whole projects. 58
• The picture he draws is a model playing its own racing game, noticing that collision detection fails, fixing it, and then fixing acceleration and detail in turn. The model judges and rewards its own work, which he says is hard to pull off because that judgment is the unreliable part. 8
• He corrects the sticker price. A bit over a dollar per million output tokens is about 15 times cheaper than Claude Opus 5.5, but in his own test DeepSeek took two to three times more tokens for the same task, so the real advantage is smaller. 8
• The weights are free to download and keep, while running them takes more hardware than he has at home. He expects that gap to close within a year as other labs distil DeepSeek down. 8
• His closing read is about the pace rather than this model: five weeks between checkpoints, with the open-weights version arriving close behind the closed frontier each time. 8

This is much more concerning than you think...

• Anthropic filed to list its Class A stock on Nasdaq under the symbol "ANT", with the number of shares and the price range still undetermined. 9
• The numbers are extraordinary by any standard. Net revenue retention is approximately 900% as of September 30, 2026; revenue for the six months ended June 30, 2026 was $16.3 billion, up 1,300% year over year. 10
• Customers: more than 11,900 spending $100,000 or more on a trailing-twelve-month basis, up from more than 950 a year earlier, with more than 1,900 at $1 million and about 250 at $10 million. 10
• One share of Class F stock, held by a company the seven co-founders control, carries 50.1% of total voting power. A separate Long-Term Benefit Trust elects and removes a majority of the board, and the filing states that holders of Class A stock cannot affect any vote. 10
• The internal numbers read the same way: over 90% of codebase contributions in September 2026 were written by Claude, revenue per employee on a trailing basis went from $1.4 million to over $6.6 million in a year, and the company had raised $141.1 billion against $112.4 billion of cash at June 30, 2026. 10

Genie 4: A new generation of world simulation

• Genie 4 generates worlds from a text prompt or an image, and you steer them while inside: change the weather, move the camera, add a person, and the simulation adapts as you prompt. 11
• It is DeepMind's first real-time world model to generate audio natively. Every world carries its own ambience, and characters hold interactive dialogue that reacts to what you do in the scene. 11
• Physics and human motion are the other advance: aerodynamics, collisions, and characters that move and respond without being puppeted. DeepMind frames the robot examples as practice before work in the physical world. 11
• Gemini does the reasoning around the simulation. Gemini 3.8 Flash acts as a guiding agent in the loop, breaking goals into steps and checking each rendered frame, while Gemini Omni anchors the world's appearance. 11
• Access starts with a research preview for trusted testers, then Google AI Ultra subscribers inside Project Genie, then early-access partners of the Gemini API and AI Studio. The waitlist is open. 11

AI Crisis in Math: Navier-Stokes, Riemann Hypothesis & Complexity Theory

• The results are the ones the field is still absorbing: Navier-Stokes, a quasi-Riemann hypothesis, the unique games conjecture, and hundreds more. Kontorovich counts the drop at about 700 problems. 12
• He describes where he was when it landed — a gala in New York, refreshing his phone at six in the evening — and what stopped him: a quasi-Riemann hypothesis covering the zeros of the zeta function and the Dirichlet L-functions. "I did not think I would see a proof of the quasi-Riemann hypothesis in my lifetime." 12
• The reason the claims can be checked at all is Lean and its mathematics library, Mathlib. Proofs are autoformalised and verified to the axioms, and Kontorovich puts the Lean formalisation of Fermat's Last Theorem at about 13 million lines, a project he had expected to take another three years. 12
• The cost to the profession is the argument's centre. Fridman reads Terence Tao's post about person-years of patient research being "casually disrupted or destroyed", and Moitra names the doorman fallacy — reducing a job to its most visible task — to describe what a mathematician does beyond proving theorems. 12
• Kontorovich's "solved axiom" is the test they leave open: if anything a human and a model could prove together could also have been proved by the model alone, what is the human for. His answer is that he does not believe the axiom yet, and the centaur chess analogy is why he worries. 12