AI designs enzymes and proposes mathematical proofs

AI is designing proteins for chemical reactions, ranking plant mutations and proposing mathematical proofs. These stories involve different kinds of results: laboratory experiments, computational predictions and manuscripts still awaiting review. That distinction determines what is actually known about each system.

The detailed AI Search review of October 11, 2026 covers the demonstrations and announcements examined below. Comparative results are reported with their sources and test conditions; availability refers to the recording date.

Google DeepMind presented AlphaProtein Novo. Its purpose is to design new enzymes for specified chemical reactions. Enzymes are proteins that accelerate chemical reactions: they participate in digestion and healing in the body, and are used industrially to make chemicals and medicines. The conventional process starts with a natural enzyme and experimentally modifies its properties. In the new approach, AI designs both a three-dimensional structure and an amino-acid sequence from scratch, then tools such as AlphaFold 3 predict folding and interaction with the target.

According to the presenter, researchers tested the designs in laboratory experiments. One produced an enzyme that helps make piperidine, a chemical used in pharmaceuticals. Another produced an enzyme that breaks down DEHP, described as a difficult-to-remove pollutant used as a plasticizer. He stresses that the research is early-stage and has not been peer-reviewed. He sees it as a step toward designing enzymes for requested reactions, including pollutant breakdown and more efficient chemical production. A GitHub repository provides running instructions. He identifies this as his most exciting scientific development of the week.

The experiments are described in the AlphaProtein Novo preprint; access to code and models lets researchers study the method but does not replace laboratory testing of a new enzyme. The generator code is licensed under Apache 2.0; the terms for weights and generated output permit their non-commercial use only by, or on behalf of, non-commercial organizations.

BOTANIC-1 is a specialist plant-DNA model used alongside Gemma 4. In the experiment described by Living Models and Google, Gemma builds the analysis while BOTANIC-1 scores variants using their surrounding DNA sequence. The model was trained on genomes from 320 plant species; the task is to distinguish a trait-changing mutation from many nearby changes.

In the melon-flower experiment, the correct mutation ranked first among 2,494 candidates. The 90% figure is the success rate across 20 runs of that task. The authors report different metrics on a broader retrospective evaluation of known mutations. The ranking helps select a hypothesis for laboratory validation; it does not establish 90% general accuracy across plants and mutations.

OpenAI published 719 mathematical manuscripts grouped into 372 result families. In its repository, the company says it used an unreleased internal model, posed approximately 4,000 research problems and generally allocated about three hours of computation per result. A family can include a principal result, supporting arguments, consequences and alternative proofs.

Verification status varies. The October 11 review mentions 162 manuscripts with Lean formalizations—computer-checkable proofs. When checked later, the repository README reported roughly 42% of top-line results formalized. Manuscripts and top-line results are different counting units. OpenAI explicitly allows that unformalized work may contain issues and says corrections will remain in its version history.

The presenter gives the release an exceptionally high assessment and explains four subjects: zeta-function zeros, Hilbert’s tenth problem, matrix multiplication and elliptic curves. His comparisons with mathematicians’ collective discoveries remain personal judgments. The number of published texts does not establish the number of independently accepted scientific results.

Prime numbers are divisible only by one and themselves. The Riemann zeta function helps study their distribution. The Riemann hypothesis describes the location of its nontrivial zeros—inputs at which the function vanishes. These inputs are complex numbers, with real and imaginary parts. Even checking a huge number of examples cannot prove the statement for every case.

OpenAI’s catalogue claims a zero-free region to the right of real part 7/8 and a separate proof for an 11/12 boundary. The review describes this as narrowing the region where zeros may occur. The full Riemann hypothesis remains outside the claimed result. The catalogue describes a proposed result; independent acceptance of its proof is a separate question.

The presenter connects interest in primes with their role in encryption, but demonstrates no new attack. The Fields Medal assessment is his own. Publishing a manuscript does not itself immediately change the security of payment cards or passwords.

The next subject is Hilbert’s tenth problem over the rational numbers. The presenter explains rational numbers as fractions, including halves, quarters and integers written with denominator one. He begins with a simple equation whose answers are 3 and −3, then moves to equations involving dozens or hundreds of unknowns. In his historical account, Hilbert’s 1900 question asked whether a general method could decide if such an equation has an integer solution. In 1970, the impossibility of a universal method was proved. Allowing fractions, he says, left a version unresolved for decades.

The presenter now attributes to OpenAI a published proof that a universal method is impossible for rational solutions too. He explains the claim as a limit on algorithmic decision-making for a particular class of problems: even a vastly more powerful computer would not supply a single algorithm that always correctly decides whether every equation in that class has a solution. Individual difficult equations can still be solved, including with AI. The substantive limit concerns a universal procedure covering all allowed cases.

The third result concerns the matrix multiplication exponent. The presenter describes a matrix as a grid of numbers and names graphics, simulation and AI as uses of matrix multiplication. Computational work grows rapidly as matrices become larger. He recalls a reorganization of the calculation introduced in 1969 and subsequent reductions in the operation count. The exponent describing how work grows with matrix size is 3 for the conventional method; he says improvements had brought it to about 2.37. The aim is to approach what he calls the theoretical minimum of 2. The new OpenAI result claims 2.25. He relates the possible significance to repeated use of matrix calculations, potentially billions of times, but supplies no measured acceleration of an operating system or application.

The fourth subject is the Birch–Swinnerton-Dyer conjecture, one of the Millennium Prize problems. The presenter considers points on elliptic curves whose coordinates are rational: some curves have finitely many, others infinitely many. Known points can be added to produce more, so in some cases a few independent starting points suffice to generate infinitely many others. In his account, work by two mathematicians in the 1960s linked the number of independent points to the behavior of an L-function associated with the curve near a particular input. According to the conjecture, a nonzero value predicts finitely many rational points. If the function vanishes there, the order of that zero—how quickly the function grows away from it—should determine the number of independent starting points needed to generate the others.

Previously, the presenter says, this relationship had been proved in special cases while the general statement remained unresolved for around 60 years. A new series of OpenAI proofs claims a broader class of elliptic curves. Combined with another related result, it reportedly establishes the full prediction for almost all curves within certain families of equations. Limits remain: the entire conjecture is not solved, some cases are outside the proofs, and the importance of the new results depends on independent verification. Ending this section, he offers his personal view that people’s capacity to read and understand AI discoveries has become the main bottleneck. He says the complete manuscript collection is on GitHub.

Vals AI described two magnetic-memory candidates found with Claude Opus 5.5 agents. The researchers seek materials with a compensated net magnetic moment but electron states separated by spin—an electron’s quantum property. That combination is of interest for closely packed memory elements and electrical readout of their state.

The first candidate, YBaMnFeO₅, is a new design. Calculations predict the desired combination of properties but require an ordered manganese–iron arrangement. Standard synthesis conditions may destroy that order. The second, KV[Cr(CN)₆], was made in 1999. That work measured magnetic ordering above room temperature; the new calculations identify a combination of semiconductor behavior and spin-sorted states.

The second material’s band gap and spin sorting have not yet been measured. An ideal dry-crystal calculation also differs from the real sample, which contains water in its pores and has a small residual magnetic moment. No working memory device has been built from these candidates. Synthesis and direct measurements are needed to test the predicted properties.

The review also covers WorldPlay 2, a generator of worlds that users can move through and interact with in real time. Comparing it with Genie 3 and Hunyuan World, the presenter highlights additional capabilities: interacting with characters and objects, and requesting effects or animations through prompts. Earlier events are compressed into memory to preserve the setting when a user leaves and returns. In a prolonged example, the presenter reports no noticeable deterioration and calls the system one of the most flexible open world generators. He says its code and instructions are already on GitHub. It is based on Alibaba’s Wan 2.2; a faster turbo version generates a world in four steps, but still occupies 69 GB because it uses the full, unquantized Wan model.

Motion Space Flow (MSFlow) generates human movement from text. Examples include a high kick with the left leg and crawling forward on hands and knees. Users can also constrain any joint’s position, specifying, for instance, where a hand or foot must be at a given moment. The rest of the animation is generated around that constraint. Code, instructions, and training and evaluation scripts have been released. The presenter puts the entire Hugging Face folder at approximately 1.1 GB; that file size does not establish VRAM requirements. The stated license is Apache 2.0.

For Kandinsky 6.0, the presenter describes video examples with intense action and complicated movement. There are two variants: Lite with 3.8 billion parameters and Pro with 29 billion. His main reservation is duration: it generates only five-second clips, whereas the MiniMax models he mentions can produce longer ones. Text-to-video and image-to-video are supported; reference-to-video is absent, although he considers it an important current standard. A separate Kandinsky 6 Super Resolution model increases the internal compressed numerical representation of frames by up to four times and can be added to the generation pipeline. According to his account, the tools are available in Diffusers, ComfyUI and other environments. Lite occupies 6.38 GB and Pro 60 GB. He personally rates its quality and flexibility below MiniMax H3, while describing it as another open alternative.

The presenter introduces Intelligent UI as a feature allowing ChatGPT to construct an interactive interface within an answer. A response explaining a bicycle includes a diagram whose parts can be selected to learn how they work. A Sunday lamb-roast planning example includes images, a cost breakdown, adjustable numbers and an interactive checklist. Another reported change concerns GPT-6: it can begin answering before completing all of its reasoning. Comparing it with the previous GPT, the presenter describes faster output. The announcement he cites says the global rollout begins with paid plans and expands to the free plan the following day; he expects the interface to be available there by the time of recording.

Amazon’s ALoDLM language model, in the presenter’s explanation, predicts multiple parts of an answer simultaneously and refines them over repeated passes. He interprets the displayed reasoning, mathematics, science and coding benchmarks as showing that the system leads similarly sized models on most tests. In one comparison, he quotes 8.5% higher throughput and 13.6% less computation. Code and models are reportedly available: the 1.7-billion-parameter version occupies around 4 GB, and the 8-billion version around 16 GB. The license is described as Creative Commons non-commercial.

Embedding Gemma 2 is described as an open Google model that transforms information into numerical representations called embeddings. Its purpose is information representation and search rather than generating text or images. Text, documents, images, video and audio are mapped into a shared representation space so another model can search them. The presenter gives 270 million parameters for text, an additional 170 million for vision and 300 million for audio, with the total below one billion. He connects the small size with efficient operation even on a Google Pixel smartphone. Stated input limits are up to 5.5 minutes of audio, 29 images or 59 video frames. He highlights strong image results among similarly sized models on the displayed charts and also shows an audio benchmark. Possible uses include search, assistants and applications working with large file collections. He reports Hugging Face availability, a size of around 1.5 GB and Apache 2.0 licensing.

We describe the application of these representations in our knowledge-retrieval project at pommeDeTerre: documents are indexed and answers are tied to retrieved passages and checkable sources.

For a website use case, our guide examines requirements for a knowledge-grounded website consultant, including source material, answer boundaries and handoff of difficult enquiries to a person.

Mistral Large 4, also called LeChonc, is presented as a mixture-of-experts model with one trillion parameters in total and 52 billion active during use. Its weights are not yet open at recording time; a release is promised by the end of the month. The presenter calls it Mistral’s best model but finds its price-performance combination uncompetitive. In his reading of the Artificial Analysis index, it trails Kimi K3, GLM 5.3, GLM 5.3 Flash and DeepSeek Flash. He also describes it as more expensive than DeepSeek 4.1 Flash and MiMo 2.6 Pro, which he judges stronger. These conclusions reflect his assessment of the comparative charts shown. He mentions access through an API, a programmatic interface through which applications can send requests to the model.

Meta’s GenIA turns a photograph or video into a complete 3D object. It uses SAM3D with a recurring check of how well the reconstructed object matches the original image. Differences guide adjustments to appearance and position. Code and local running instructions are available; the review notes a Creative Commons non-commercial license.

FlashDexRetarget transfers demonstrations of human hand movement to robotic hands. One learned control policy covers many movements, without separately training a system for each demonstration. It also uses several upcoming frames to account for where the hand and object are moving. The presenter reads the displayed tests as showing better average results than earlier methods. Code and instructions are available. The review reports around 110 GB of GPU memory for training. The reported evaluations use XHand and Sharpa Wave Hand robotic hands.

The open Iris 3B model has three billion parameters and operates directly in pixel space without a VAE, a component that compresses an image into an internal numerical representation. The presenter describes this as theoretically capable of improving precision and quality. Reported examples include photorealism, different artistic styles and large text. Text-to-image and image-to-image are supported; semantic editing of the kind offered by Nano Banana is unavailable. He judges the displayed photographic quality below the other generators in the comparison shown, while highlighting further training for particular tasks. Examples include depth estimation and upscaling. He reports available code, data and training scripts, with Apache 2.0 model licensing. The base model files occupy 12 GB; VRAM requirements have not been separately verified.

Terra trains simulated human bodies to move using bones and muscles. The presenter contrasts direct joint control with simulating muscle contractions that pull on bones to create movement. Training used more than nine hours of motion data and 18 demonstrated movements across different terrain. Afterward, the general muscle model reportedly handles slopes, steps and other difficult surfaces. Additional specialized, fine-tuned policies allow jumping and rolling. He reports a GitHub repository, local running instructions and Apache 2.0 licensing.

CoDimRecon converts ordinary environment photographs into an editable 3D scene ready for physics simulation, including deformable objects’ physical properties. It first reconstructs the overall geometry, then represents objects differently: a cable as a flexible 3D curve, paper or a plastic bag as a flexible surface, and a cushion as a deformable solid. Robot interactions can test behavior, and the scene can be revised if results are unsatisfactory. The presenter calls it the only tool combining all these capabilities. Code is merely promised at recording time.

DistScene reconstructs a complete scene from one image, including the environment and individual objects. In the described process, these are first generated together in a shared 3D coordinate system. Each object is then extracted, its geometry refined individually, and returned to its correct position. The presenter interprets displayed comparisons as the best results in geometry, consistency and quality. At the time of recording, release of the model and dataset was still planned.

The presenter considers Nano Banana 2.1 a relatively small update. He attributes to its developers improved generation and editing under complex instructions, with multiple characters and text inside images. He mentions infographics, diagrams and precise natural-language edits, up to 14 reference images, improved realism and text rendering, and resolutions up to 4K. His own initial tests leave a less favorable impression: he judges quality noticeably below GPT image 2.5. He nevertheless says registration on Google Flow provides a few dozen free generations per day.

The final substantive section concerns Anthropic’s release of Claude Haiku 5.5. The presenter describes Haiku as the smallest and fastest variant, and Opus as the largest and most intellectually capable. He reads the displayed knowledge-work, agentic-coding and computer-use tests as showing substantial improvement over the previous Haiku and advantages over GPT-6 Luna in most comparisons. Relative to Sonnet 5.5, he says the new Haiku is slightly less capable but considerably faster and cheaper. Artificial Analysis places it behind Kimi K3, Step 5 and GLM 5.3. The quoted cost is around 21 cents per task, below GLM 5.3 and DeepSeek Flash, but approximately three times GPT-6 Luna’s cost. For the combination of price and capability, he prefers Xiaomi MiMo 2.6 and also rates GPT-6 Luna highly. According to his account, Haiku 5.5 is available even on the free plan and can be selected for the fastest response.