Gemini 4, persistent agents and the coded Marmont letter

OpenAI introduced persistent Dots agents, Google presented Gemini 4 Argon with restricted access, and small speech-recognition models reached sizes measured in tens of megabytes. Image editing and scene reconstruction sit alongside tasks with checkable outcomes: a mathematical construction, a robot’s contact with the ground or a deciphered letter.

The detailed AI Search review of October 4, 2026 examines these developments. The comparisons and availability details below refer to the recording date.

OpenAI Dev Day introduced Dots, a persistent agent with its own cloud computer and browser. The presenter describes connections to thousands of apps through plugins and continuous work toward a goal even while the user is not chatting. His example is monitoring customer feedback, fixing bugs, testing changes and preparing pull requests. A Dots agent can receive several projects without separate conversation threads. Communication is described through ChatGPT, Slack and Teams, including from a phone, with messaging and calls mentioned. According to the announcement he cites, rollout covers Pro and Business Premium in eligible markets, excluding the EU. GPT-6 Astra is named as the underlying model. He compares the concept with OpenClaw, Meta Muse and Grokbot. The official announcement separately describes proactive research as a read-only mode. Continuous execution of assigned tasks does not authorize the agent to make unsolicited changes in connected apps.

The presenter says two separate open alternatives, both called OpenDOTS, appeared within hours of the announcement. The first can be self-hosted and connected to a chosen language model, including a local model with suitable hardware. It is a framework requiring the user’s own LLM. The second likewise supports self-hosting and assigning tasks to agents. For that project he specifically lists text messages, calls and Slack interactions. He presents both as open implementations of the same general persistent-agent arrangement without conducting a comparative test.

Connecting business systems and configuring permissions remain application work. Our AI agent development service defines these boundaries through agreed integrations, data access and permitted actions. Dots capabilities alone do not determine permissions inside a particular company.

Another announcement is GPT-6.1 Sol, updating GPT-6 Sol, which the presenter says appeared a week earlier. He relays OpenAI’s claim of near-Astra intelligence at one fifth of its standard input- and output-token prices. Completed-task cost also depends on token usage. The comparisons shown cover agentic coding, AutomationBench, interface operation and scientific research. In Artificial Analysis’s index, the model is one point behind GPT-6 Astra with substantially lower task costs. He calls it the most cost-efficient frontier model, his assessment of those charts. Announced availability covers paid GPT Work and Codex plans, with ordinary chat still excluded at the time of the recording.

Alongside the cheaper model, he highlights a subscription-limit change. For the $200 Pro plan, usage relative to Plus was set to drop under the change announced then from twenty times to ten times. He describes this as halving the previous relative allowance. He also mentions a new $500 plan offering more usage. The review gives no specific allowance for that plan.

The presenter describes Gemini 4 Argon as Google’s new flagship. In the displayed coding, knowledge-work, science, mathematics and computer-use tests, he highlights generally higher scores than Astra and Claude models. He names first-place results for long-horizon DeepSWE coding, Val’s index, financial research and extended legal work. In the separate Artificial Analysis ranking, however, it ties for second with GPT-6 Astra and Claude Fable, several points behind Opus 5.5. He considers its lower cost relative to similarly capable models an important advantage.

He particularly emphasizes the claimed ability to output up to one million tokens in a single response. He explicitly distinguishes this from input context size and compares it with other frontier models’ 64,000–128,000 output tokens. He approximates one million tokens as 700,000 words. A novel, several novels or an entire codebase in one response serve as illustrations of that volume; he does not execute such a task. He separately asserts that all frontier models already have million-token input contexts.

For the Omniscience benchmark, he quotes a 15% hallucination rate for Gemini, around 50% for GPT-6 and roughly 60–70% for Claude. He explains the condition: the benchmark examines behavior when a model does not know an answer, including wrong or partial answers versus admitting uncertainty. He interprets the result as a greater willingness to acknowledge missing knowledge rather than a correct-answer percentage. At the recording date, access is limited to trusted cyber defenders in a private program. A broader rollout was then planned to begin with paid API customers and Google AI Ultra subscribers. He postpones his own public-model comparison until release.

InSpatio-World 1.5 generates a video moving through a scene from an image and a specified camera path. Supported inputs include a single image, multiple images for better consistency, a panorama or a complete video. With video input, the system also captures the original motion while the camera moves through the scene. It can freeze the action to create a bullet-time effect. The review reports that code and running instructions have already been released. The review gives a file size below 6 GB.

NVIDIA’s SoL-Refiner upscales video to 4K in one step. The presenter emphasizes compatibility with different generators, with examples including MiniMax H3, WAN and NVIDIA Cosmos. In the comparisons he describes, the result is sharper but changes individual details. He specifically notes altered elements of a barn: increased detail comes with departures from the original image. Fidelity is his main objection to the demonstrated approach. He says code and local running instructions are available.

The speech-recognition model Whistle occupies approximately 16.9 MB. The presenter says it runs on a CPU without a GPU or dependencies. He dictates a passage and shows its transcription. It supports up to 30 seconds of audio per pass and automatic language detection. Word-level start and end timestamps are also described, although that feature is not shown in the demonstration. The presented comparisons indicate decoding roughly six times faster than WhisperBase and lower error rates on some benchmarks. Installation uses pip, and models are hosted on Hugging Face. He sees its small size as suitable for phones and other devices without GPUs.

The next system, Phonon 2, is below 900 MB. Its displayed page claims that an hour of audio can be converted to text in around 20 seconds. The demonstration also shows the presenter’s own input and the resulting transcription. In the table he shows, it has the second-lowest average error rate among the compared systems, including substantially larger ones. He describes it as open source, with downloadable models and free Hugging Face Spaces for trials. Both small systems are presented in the context of speech recognition and converting audio into text.

Recognition is one stage of processing a recording. In GistiQ for meeting recordings, the transcript supports search and work with the conversation’s content. This is a separate studio product; the results in this review do not describe its quality or establish that it uses these models.

The presenter compares Anthropic’s new Sonnet 5.5 with its predecessor and the larger Opus 5.5. In the agentic coding and knowledge-work tests shown, he highlights substantial improvements and scores close to Opus. In Artificial Analysis’s intelligence index, Sonnet is second, ahead of GPT-6 Astra. He separately considers cost per completed task: the displayed figures are higher than for Claude Fable and Opus. He associates this with high output-token consumption and calls Sonnet the most expensive model in that comparison. His preference for Opus or Fable for intelligence, and GPT-6.1 Sol or Astra for value, is his assessment. He also says Sonnet 5.5 is now the default model on the free plan.

Comfy Agent is built into ComfyUI to create workflows from user descriptions. It plans steps, creates and connects nodes, and runs the resulting workflow. It can also edit an existing workflow. Error correction is described more cautiously: users can ask the agent to try fixing an error automatically. One example request is generating a MiniMax video with an appropriate workflow selected automatically. At the time of recording, the feature is available in paid Comfy Cloud, with Comfy Desktop support promised in a few weeks.

PixelUMM is presented as a research architecture that understands and generates images and videos directly in pixel space. The presenter contrasts it with systems that first generate a compact internal representation and then convert it into a visible image. He explains that this separate conversion stage is removed. He considers the shown samples weaker in quality and consistency than frontier models, while emphasizing the project’s research status. Code and models have been released with local running instructions. Most source files are described as Apache 2.0 licensed.

The developers of Ideogram 4.5 call it their most precise image editor. In the examples described, text instructions change selected elements while preserving the pixels in other areas. Across repeated edits the presenter notes less accumulated distortion than in the shown GPT-Image and NanoBanana outputs. As in version 4, users can specify a rectangular region to constrain changes. At the recording date, an online platform is available, while an open release for local use was only planned. His expectation that it will become the best local editor is his opinion.

FLUX.3 Image also supports rectangles with separate descriptions for different regions of an image. Ordinary text prompts work as well; the presenter describes realistic photographs and complex posters with text. Up to ten reference images can be supplied. Editing functions include adding, removing and replacing objects, changing colors, upscaling and colorizing black-and-white photographs. At the recording date, the model is closed and available through an online platform. He qualifies the promise of open weights by saying a similar promise for FLUX.3 video had remained unfulfilled for several months.

ByteDance presents DMAD, Distribution Matching as Adversarial Distillation. The presenter explains it as reducing video generation from the usual 20–30 steps to four. A base model produces training outputs, and a new model learns to reproduce them in fewer steps. Discriminator components compare the two models’ outputs and provide an improvement signal. He says the approach applies across different video and image models. Code and instructions are released. The described setup requires the base MiniMax model, with additional LoRAs of 1.4 GB each.

ByteDance’s second method, PDMD, also enables four-step generation. The presenter rates the displayed motion, detail and consistency above other distillation methods. His explanation concerns errors in the evaluating model: inaccuracies in its signal can transfer into the model being trained. PDMD, he says, identifies the signal component arising from that error and projects it away. The displayed comparison metrics are better on average. Released code accompanies two options: a complete modified MiniMax model of 66 GB and a roughly 1.4 GB LoRA usable with a smaller quantized base model.

A stellarator confines plasma with magnetic fields of complex three-dimensional shape. The mathematical question concerns smooth equilibrium configurations with varying pressure and no axial symmetry. Equilibrium means satisfying force-balance equations; stability under perturbations requires separate analysis.

The Counterexamples to Grad’s conjecture repository publishes a Lean formalization of its mathematical result. A separate project, Analytic 3D MHD equilibria, contains explicit analytic solutions, checking programs and the original GPT-6 Astra Pro conversation used to find the solutions. These works present mathematical constructions; they include no tests of an operating fusion reactor.

The Analytic 3D MHD equilibria materials allow checks that the given fields and pressure satisfy the stated equations. The presenter connects this story to earlier reports of fluid-equation singularities and expects further AI results. That assessment does not extend the specific works’ proven conditions to every plasma or fluid problem.

Tactile Step adds pressure information from under the feet to walking control. Successfully traversing terrain can still involve hard impacts, edge landings and unstable contact. Researchers equipped a Unitree robot with pressure-sensing insoles and further trained its controller to account for each foot’s contact location and force. The demonstrations described include stairs, slopes, platforms and level ground. Reported results include up to nearly 50% lower peak touchdown force, reduced impact noise and up to 24% greater contact area. Those figures are presented as the system authors’ results.

Another Unitree robot, the G1, is shown playing badminton autonomously against a person. The work is attributed to Tsinghua and other universities. The demonstration shows forehands, backhands, jumping returns and several consecutive exchanges. The presenter connects the difficulty with shuttle speed and the short time available to coordinate running, leg movements, torso rotation and racket swings. Because high-quality human motion recordings were scarce, the researchers used a limited set of real hitting motions and automatically generated synthetic variations. That expanded dataset was used to train the robot.

ElevenLabs V4 is a speech-synthesis model its developers call their most emotive voice model. Users can insert metatags specifying delivery, tone and sound effects. The review plays several examples: a threatening address, emotional family speech, and sports commentators’ conversation shifting into excited play-by-play. They illustrate changes of intonation and delivery within a passage. The presenter says the model is available on the free plan with an initial 10,000 credits. He attributes its first-place speech-synthesis standing to a displayed leaderboard rather than a systematic comparison of his own.

In the story about a coded letter to Marmont, the presenter says GPT-6 Astra helped solve a cipher that had remained undeciphered for about 217 years. Carter Church started from a low-resolution image. The model split it into rows, recognized roughly 1,300 handwritten cipher units and grouped differently drawn versions of the same symbol. A previously published table supplied values for only 33 letters and covered about a third of the text. The model then built a solver testing mappings against plausible French. After several rounds it reconstructed the remaining key, including symbols for whole words. The process took around six hours. The letter is described as a military briefing to General Marmont on troop positions before Austria’s war with France.

PAMI generates a 3D animation of a person interacting with an object from a text prompt and that object. Examples include picking it up, moving it and rotating it. The presenter highlights more natural holding than in comparisons where hands become disconnected from the object. He explains that object movement combines signals from hands, feet and hips, after which a refiner adjusts the preliminary interaction to correct contacts. A GitHub repository was available at the recording date, while a model release was planned. MIT licensing is mentioned.

Point2Part divides an object using points placed by the user on the desired parts. Its input is a single image or a 3D mesh. The demonstrations show separate components that can be spread apart for inspection. The system decomposes the full shape using all specified points. In the presenter’s explanation, the parts therefore neither overlap nor leave gaps and reassemble into the complete original object. He cites part-quality benchmarks that outperform similar tools on average. A repository is public, while models and training code were still promised at the recording date.

Allen AI’s AstaBrief turns a prompt and relevant scientific papers into a cited research report in one pass. Its stated base is Qwen3-8B, with eight billion parameters. In the displayed report-quality evaluations, the specialized system substantially improves on the base model and approaches Dr. Tulu. Models and datasets are released. The full version is around 32 GB; the presenter cautions that local use still requires a high-end GPU. He presents it as an option for scientific-literature reporting when suitable hardware is available.

The same institute’s OLMo Core 3 is presented as a training system for mixture-of-experts models. Only some expert components activate during operation, reducing computation relative to using every parameter. The presenter explains the data-transfer bottleneck across hundreds of GPUs. In the described solution, experts stay on assigned GPUs while computational work is sent to them. Models and training data are additionally distributed across devices to save memory, and smaller calculations are grouped for more efficient chip utilization. Combined training speedup over the previous system is reported as roughly 2.7 times. A technical report, code and documentation are released.

iQuest Q1 targets long software tasks in command-line environments and tools such as Claude Code and Codex. Listed actions include inspecting repositories, running commands, debugging, using tools and iterating until completion. It is a 320-billion-parameter mixture-of-experts model with 15 billion active parameters. The presenter favorably assesses the displayed benchmarks in relation to its size. Code is released, and the repository contains download and running instructions.

The final model AREX-2 has 27 billion parameters and is based on Qwen3.8-27B. Its training focuses on iterative improvement: proposing a solution, testing or measuring it, examining feedback and trying again. Machine-learning and algorithmic-programming tasks with checkable results received particular attention. The presenter reports performance comparable with larger models and highlights results relative to parameter count, mentioning Kimi K3 and GPT-5.6 Sol as well as deep-research tests. The model is downloadable from Hugging Face. Its complete version occupies 55 GB, a qualification he adds after describing it as small.