Today’s editorial: We discuss why DeepSeek Harness is even bigger than it seems |
|
| | | | We recommend: | 👨🔧 MongoDB Agent Skills: Now Live | | Official MongoDB Agent Skills are live! | Bring MongoDB best practices into Claude Code, Cursor, Gemini CLI, and VS Code. | |
| |
|
| | | | This Week in Turing Post / Summer schedule: | Friday / New: I’m pondering an idea for a new type of media that I’d like to share with you. Sunday / Library: Chinese LLMs in 2026: DeepSeek, Qwen3, Kimi K2 and More
|
| |
|
| | | | You all are missing out | When DeepSeek released DeepSeek-R1 in January 2025, it blew everyone’s mind. R1 came surprisingly close to OpenAI’s o1 on several reasoning, math, and coding benchmarks, and DeepSeek released the weights under an MIT license along with a paper explaining much of its RL process. V3 had already shown that a Chinese lab could build a very capable model efficiently. R1 went further, showing that even advanced reasoning performance was no longer exclusive to a handful of frontier labs. They still had enormous advantages in compute, distribution, talent, and proprietary data. But the model itself was starting to look like a less reliable moat. | Now, DeepSeek Harness may be doing something similar one layer higher. | It has received enormous attention on GitHub, with more than 149,000 stars at the time of publication, but most of the discussion still treats it as another open-source coding assistant. I think there is more to it. DeepSeek has published its version of the layer that turns a model into an agent: tools, memory, execution loop, sandbox, storage, scheduling, and interface. | The slogan is “Everything is a Plugin.” At first, that sounds like a fairly standard software idea. But sometimes a simple abstraction becomes much more interesting when it is applied to the right layer of the stack. | When “everything is” becomes an architecture | Unix with “everything is a file” made devices and streams accessible through file-like interfaces. Smalltalk with “everything is an object” organized its environment around objects communicating through messages. The Web (the hourglass model: everything over IP, IP over everything) separated documents from the machines serving them through common protocols. | In each case, many of the underlying components already existed. The advance came from giving them a common interface that allowed previously separate pieces to work together. | Alan Kay described Smalltalk as a response to rapidly falling computing costs. At PARC, cheaper personal computers changed the assumptions about what software could look like. The architecture followed from that. | Harness starts from a similar change, this time in the cost of producing software. What should an agent look like when writing another piece of software becomes cheap enough to do in the middle of a task? | Most assistants today are still fixed applications connected to a set of tools. Harness treats almost every application-level capability as a plugin. And in creator mode, the agent can go further: if it needs something it does not have, it can write a temporary plugin, ask for permission to load it, use it, and remove it afterward. | So some capabilities no longer have to be built into the application in advance. They can be created when the agent needs them. | Why the paper is more important than the demo | Code generation lowers the cost of adding software, while the cost of managing everything that has already been added continues to grow. | A generated tool can register commands, alter shared state, depend on services, or leave other components expecting resources that no longer exist. An agent that continuously writes extensions without managing their lifecycles becomes less reliable with every improvement. | Harness is built on Cordis, which combines two ideas. | First, effects should be reversible: a component records how its changes can be withdrawn. Second, dependencies should be reactive: if a required service disappears, dependent components deactivate before it is removed.
| The accompanying paper, “A Programming Paradigm for Spatiotemporal Composability,” tries to formalize these ideas. Under a set of assumptions about reversibility, dependencies, and independent effects, the authors show that a system that changes dynamically can end up in the same observable state as one built from scratch in that configuration. | This resembles what Edgar Codd did for databases in 1970. His relational model converted data independence from an engineering preference into a formal abstraction. SQL later diluted parts of the original vision, but the abstraction survived and reorganized the industry. | It is far too early to say whether Cordis will play anything like that role for agent runtimes. But I think it is asking the right question. If agents are going to create their own capabilities, we need a good way to add those capabilities, combine them with everything already running, and remove them again without breaking the system. | What’s interesting is that Cordis did not start with agents. It grew out of Koishi, a cross-platform chatbot framework created by Shigma, who now works at DeepSeek. Over four years and thousands of community plugins, Koishi ran into many of the practical problems that come with dynamic software: capabilities could appear or disappear, plugins depended on one another, and the system still had to keep running as its composition changed. | It also makes sense that this architecture emerged from chat. In China, messaging platforms have long combined communication with payments, commerce, public services, entertainment, and work. WeChat alone has around 1.4 billion users, while Koishi supports platforms including QQ, Telegram, and Discord. | Chatbots were already dealing with some of the problems agents face today: continuous events, persistent state, external services, and independently developed tools. In that sense, chatbot frameworks already had some characteristics of agent runtimes. Once language models became capable of choosing and creating tools themselves, the plugin-lifecycle problems Koishi had been dealing with became much more relevant to AI systems. | The second moat | After the first DeepSeek moment, the conversation about defensibility started moving up the stack. If strong models become cheaper and the differences between them continue to narrow, perhaps the valuable part is the harness around the model: tools, context management, permissions, execution, memory, and all the workflow design accumulated over time. And for a little while, it felt like this might become the moat. | That’s what makes me so excited about Harness. DeepSeek has now released an implementation of that layer under an MIT license. It is also surprisingly easy to try. With Node.js installed, one command launches the Web interface locally (or, you can ask your other coding agent to set it up for you). And, at least in my experience, it is blazingly fast and cheap if you use the latest DeepSeek model. | | The model is not tightly coupled to the harness either. DeepSeek is the obvious default, but the model provider is itself a plugin. You can connect other models through compatible adapters, although some configuration may be required. So you don’t have to get the model, tools, and runtime from the same company. | This feels familiar. With R1, DeepSeek showed that strong models might not remain concentrated among a handful of labs. Now it is applying a similar idea to the layer above the model, right when everyone is starting to treat that layer as the next moat. | There are plenty of caveats. Harness is new, and some of its guarantees have obvious limits. Removing a plugin cannot unsend an email, for example. Reversing changes still depends on plugin authors describing their effects correctly. And DeepSeek has not shown that Harness is better than the leading coding agents. | But that is not really why I find it interesting. DeepSeek’s first big release made the model moat look less secure. Now it is opening up the layer everyone had started looking to as the next moat. | Everyone is watching to see which coding agent wins. DeepSeek is making me wonder how much of the coding agent needs to be a proprietary product in the first place. | A huge thank you to my partner, Will Schenk, for the inspiration and the Smalltalk comparison. | If any of those thoughts resonate with you – share them across your social networks. Let’s keep the conversation going. | |
| |
|
| | | | 📹 Why DeepSeek Harness Is The End Of Coding Agents as We Know Them. Check it out → |  | Why DeepSeek Harness Is The End Of Coding Agents as We Know Them |
|
|
| |
|
| | | | | | We are reading / watching | |
| |
|
| | | | | | News from the usual suspects ™ | Models sped up, APIs got pricier, and routing became a business. Such was the week. | OpenAI: frontier intelligence gets an express lane | | Anthropic: every answer leaves a trace | | NVIDIA: routing the models, financing the buildings | NVIDIA released Nemotron 3.5 Lightning, a 30B open Mixture-of-Experts model for high-volume agents, alongside NeMo Switchyard, an open router that selects models by quality, latency, and cost. NVIDIA says Switchyard can preserve near-frontier accuracy at roughly one-third the cost of using Opus 4.8 alone. (NVIDIA Blog) Six days later, NVIDIA offered up to $105B in guarantees for OpenAI’s 8GW Ohio campus, invested $1.5B in its developer, and became the site’s exclusive chip supplier. NVIDIA now routes the workloads and helps finance the buildings. (Jensen’s X)
| Google: Flash again, sign language at last | Google released Gemini 3.7 Flash for coding and agents at an introductory $0.75 per million input tokens and $3.75 per million output tokens, only three weeks after 3.6 Flash. (blog.google) DeepMind also put its multilingual SL2T sign-language model into Gboard and Live Transcribe on Pixel 11, beginning with American Sign Language to English. (Google DeepMind)
| SpaceXAI: the agent receives a computer | SpaceXAI launched Grok Bot, persistent agents with their own cloud computers that can work across apps, inboxes, and websites. It followed with Grok 4.6, built for long-running coding and knowledge work, which shipped first in Cursor and then entered GitHub Copilot. Owning the editor does not prevent distribution through the incumbent. (SpaceXAI)
| Mistral: sovereignty gets a reservation system | Mistral launched regional inference in Europe and the U.S., an SLA-backed Priority Tier, and support for third-party open models beginning with GLM-5.2. It is also asking European customers to make multi-year commitments that could support up to 1GW of Mistral-operated compute by 2030. Sovereign AI now comes with data residency, reserved capacity, and advance purchasing. (Mistral AI)
| DeepSeek: the low-price era gets peak hours | DeepSeek released V4 Pro and raised V4 API prices by between 50% and 1,100%, depending on the model, token type, and time of use. At peak rates, V4 Pro costs $1.32 per million input tokens and $3.96 per million output tokens. DeepSeek remains inexpensive beside U.S. frontier models, but the race to the bottom has acquired surge pricing. (DeepSeek)
| Z.ai: controlled access reaches open weights | Z.ai released GLM-5.3 and reported an 84.5% score on CyberGym, slightly above Mythos 5 for vulnerability discovery, while trailing it sharply on exploit development. The company delayed the open-weight release for two weeks and will restrict its most sensitive cyber functions to verified users. These are company-reported benchmark results, but the policy shift is real: Chinese open-model labs are adopting access controls associated with closed American systems. (Z.ai)
| IBM: open on Tuesday, closed on Thursday | IBM signed a $240M agreement with Together AI for a 2,000-chip NVIDIA B300 cluster dedicated to serving open models on IBM Cloud. Two days later, IBM partnered with OpenAI to embed GPT-5.6, Codex, and ChatGPT Work into IBM Consulting. IBM chose both sides, which is often the correct enterprise answer. (Reuters)
| Stripe: OpenRouter sells for more than $7B | Bloomberg reported that Stripe finalized an agreement to acquire OpenRouter for more than $7B, only months after the model-routing company raised money at a reported $1.3B valuation. The final price could still change. The company that routes payments is buying the company that routes models. (Bloomberg Law)
|
| |
|
| | | | Other interesting models | Google: Gemini 3.7 Flash becomes the new workhorse Google released Gemini 3.7 Flash on August 13 for coding, agents, complex documents, and business automation. It improves substantially over 3.6 Flash and launched at half that model’s original price per million tokens. →read the announcement Alibaba: Qwen3.8 opens at both ends of scale Alibaba released the open weights for Qwen3.8-2.4T-A95B on August 12 and Qwen3.8-27B on August 14. The first brings a Qwen-Max-class model into the open; the second packages native multimodality, controllable reasoning, and agent execution into a dense model that is far easier to deploy. →view the models DeepSeek: V4-Pro becomes the production model DeepSeek released V4-Pro on August 13 with stronger agent execution, configurable reasoning effort, and native support for the OpenAI Responses API used by Codex-style systems. It is available through DeepSeek’s app, web interface, and API. →read the announcement xAI: Grok 4.6 targets long-running agents xAI released Grok 4.6 on August 12 with a focus on sustained research, repository-scale coding, and interactive visual work. The model was trained across agentic environments and shows more self-testing and verification during extended trajectories. →read the announcement Z.ai: GLM-5.3 combines coding with unexpected cyber capability Z.ai announced GLM-5.3 on August 14 as a major coding and agent upgrade produced largely through expanded post-training. Its cybersecurity capabilities developed faster than expected, leading Z.ai to delay the open-weight release for approximately two weeks while it completes safety testing and hardening. →read the announcement Meta: Muse Glimmer brings agents onto one GPU Meta released Muse Glimmer on August 10, a 30-billion-parameter open-weight model optimized for local coding, tool use, multimodal work, and always-on agents. A quantized version fits within roughly 20GB, allowing the complete agent model to run on a single consumer GPU. →read the announcement NVIDIA: Nemotron 3.5 Lightning becomes the execution model NVIDIA released Nemotron 3.5 Lightning on August 11, a sparse 30-billion-parameter model that activates three billion parameters per token. It is designed as a fast execution layer for long-running agents, with weights, training data, and recipes released openly. →read the announcement OpenAI: GPT-5.6-Cyber becomes a restricted specialist OpenAI released GPT-5.6-Cyber on August 10 through its controlled Daybreak Red program. Built on GPT-5.6 Sol, it is specialized for zero-day discovery, exploit validation, and other advanced defensive-security work that general-purpose models frequently refuse. →read the announcement Google DeepMind: sign language reaches a production translation model Google DeepMind introduced its multilingual sign-language-to-text model on August 12. It now powers sign-to-text dictation in Gboard and Live Transcribe on Pixel 11, beginning with American Sign Language to English, moving sign-language AI from a research demonstration into a consumer product. →read the announcement LTX: open video becomes production-ready LTX released LTX-2.5 on August 11 with native multishot generation, persistent continuity across cuts, a diffusion video decoder, cinema-grade EXR support, and a raw pretrained checkpoint for deeper customization. The release pushes an open video model toward professional production workflows rather than isolated short clips. →view the release MiniMax: Music 3.0 generates the complete song MiniMax released the open weights for Music 3.0 on August 13. Given a concept and optional lyrics, it composes, arranges, performs, and produces a song lasting up to five minutes within one generation, using separate global and local models for musical structure and acoustic detail. →read the announcement Liquid AI: vision and tool use move to the edge Liquid AI released LFM2.5-VL-3B on August 12, an open-weight vision-language model for screens, documents, grounding, multiple images, and function calling. It is designed for low-latency local deployment and can run within approximately 3GB of memory. →read the announcement Cohere: native-resolution vision shrinks to 2.4 billion parameters Cohere Labs released North Micro Vision on August 12 under Apache 2.0. The model preserves the original resolution and aspect ratio of documents, charts, forms, and screenshots, providing a compact foundation for specialized visual systems. →read the announcement Upstage: Solar Pro 4 focuses on finished enterprise work Upstage released Solar Pro 4 on August 11 with a 512K context window, up to 128K output tokens, and support for English, Korean, and Japanese. It is trained for long documents, terminal work, tool use, and multi-stage assignments where values must remain consistent across spreadsheets, reports, and slides. →read the announcement Writer: Palmyra X6 optimizes cost per completed task Writer released Palmyra X6 on August 13 as its new enterprise flagship, post-trained from GLM-5.2 for grounding, tool use, delegation, and long-running marketing and revenue workflows. It is designed specifically for Writer’s agent harness and can work toward one objective for up to eight hours. →read the announcement Sakana AI: orchestration becomes a model Sakana AI released Fugu on August 13 as an orchestrator for complex, multi-step work and upgraded Namazu to a new generation with stronger Japanese and agent capabilities. The models are integrated with sandboxed code execution and can produce complete files and visual artifacts inside Sakana Chat. →read the announcement Important model papers first published this weekMotif Technologies: extreme sparsity becomes the architecture Motif 3, published August 10, contains 314 billion total parameters while activating 13.2 billion per token through eight of 384 experts. Its post-training consolidates several specialist teachers through on-policy distillation, making fine-grained sparsity and capability merging the central model-design bet. →read the paper Pathway: reasoning stays latent BDH-CQ, published August 10, is a 150-million-parameter recurrent model whose internal memory updates from examples supplied during inference. It then solves problems through iterative latent computation without producing a textual chain of thought, offering a compact alternative to long verbal reasoning traces. →read the paper Luna-TTS: speech generation becomes parallel Luna-TTS, published August 12, is a 0.6-billion-parameter diffusion-based speech family trained across Chinese, English, Japanese, and Korean. One version generates the complete acoustic-token grid in parallel, while the real-time version works block by block for streaming speech, voice cloning, editing, and emotional control. →read the paper Alibaba and LiblibAI: human animation becomes continuous LiveAnimate, published August 12, converts a 14-billion-parameter video diffusion model into a causal streaming generator. It retrieves earlier pose-relevant context from a bounded cache, allowing real-time human animation to continue for several minutes without memory use growing with the video. →read the paper Shanghai AI Lab Intern-S2-Preview, published August 13, is a 397-billion-parameter scientific agentic model trained across scientific documents, images, time series, tools, reinforcement learning, and long-horizon tasks. It also experiments with a separate parametric memory module that adds scientific specialization without changing the frozen foundation model. →read the paper Intern-S2-Mobius, published August 14, separates a globally shared feed-forward memory from attention-based reasoners that repeatedly query it. The architecture is designed to store knowledge once while allowing reasoning computation to recur, with the larger adaptation reporting nearly four-times faster inference. →read the paper
DreamX and Alaya: world models become controllable and persistent DreamX-Phi, published August 13, predicts robotic futures from a prescribed trajectory and adds geometric, depth, and object-level constraints so that visually convincing rollouts still follow the commanded action. Alaya-EVOKE stores persistent scene geometry outside the active context, retrieving only the state required for the current camera view as a generated world expands. →read DreamX-Phi →read Alaya-EVOKE DFM: permissible data becomes a model constraint Mimir v1, published August 13, is an open one-billion-parameter hierarchical reasoning model trained from scratch with permissible post-training datasets. Its importance is the attempt to treat data provenance as part of the model architecture and release strategy rather than a disclosure added afterward. →read the paper
|
| |
|
| | | | Research | Trends we see looking at every paper related to AI and ML published last week: | Agents move their continuity into harnesses, worlds, and checkpoints AI science shifts from writing papers to running recoverable experiments Safety moves to runtime contracts and internal interventions World models stop hiding state and dynamics inside pixels Computation becomes selective, parallel, and structurally cheaper
| Agents move their continuity into harnesses, worlds, and checkpoints | 🌟 AI4AI at Test-Time: Strong-to-Weak Capability Transfer via Harnesses
Transfers capabilities during inference by having a stronger model construct deterministic scaffolds for a weaker one, nearly doubling average performance without changing the weaker model’s weights. →read the paper 🌟 Persistent Recursive Worlds Enable Autonomous Software Evolution
Makes the software project and its accepted history persistent while finite-lived agents recursively delegate and propose changes, moving continuity from the agent into the environment. →read the paper 🌟 AgentRewind: Recoverable Execution for Long-Horizon LLM Agents
Records aligned checkpoints of model context and environment state so agents can roll back, preserve lessons from failed attempts, and resume from recoverable points. →read the paper 🌟 Demystifying Agent Skills: Why They Work-Until They Don’t
Shows that skills mainly stabilize procedures rather than add knowledge, while retrieval precision falls and brittle assumptions accumulate as skill libraries grow. →read the paper 🌟Envs-FORGE: Frontier-Optimized Reward-Grounded Environment Synthesis for Agent RL
Uses verifier rewards to rewrite tasks, tests, fixtures, and containers together, then admits only executable environments that pass validation into agent reinforcement learning. →read the paper Handover of In-Context Learning State Across Session Boundaries
Formalizes what a session handover must preserve, separating exact recovery of earlier material from preservation of the continuation task’s predictive behavior. →read the paper AutoDesign: Meta-Harness Optimization for Long-Horizon Agentic Design
Uses a meta-harness to revise a design agent’s workflow from rollout feedback, allowing the scaffold around the model to become an optimization target. →read the paper
| AI science shifts from writing papers to running recoverable experiments | Spark-to-Paper: End-to-End Research Paper Generation as a Composable Skill
Builds research as a set of composable skills that retrieve literature, run experiments, revise claims against evidence, generate editable figures, and apply deterministic integrity checks. →read the paper Mechanist: AI as a Scientific Instrument for Discovering the Mechanisms of Intelligence
Automates mechanistic discovery by generating hypotheses, running causal interventions, and turning resulting explanations into model-control experiments. →read the paper OmniScientist: An Omni-Modal Omni-Discipline AI Scientist
Reasons over raw multimodal scientific evidence while enforcing novelty, statistical validity, provenance, and numerical traceability in executable code. →read the paper ScienceFlow: A Long-Horizon Agent for ML Research, Scientific Discovery and Beyond
Organizes autoresearch into recoverable executable states, re-anchors work from live or archived progress, and allocates compute according to validated evidence. →read the paper
| Safety moves to runtime contracts and internal interventions | 🌟 Stealing Reasoning Traces from Proprietary LLM APIs
Shows that encrypted reasoning objects can be replayed across sessions, users, and models to recover hidden traces, transport private information, or carry invisible instructions. →read the paper 🌟 Agent Safety Should Be a Runtime Contract
Reframes agent safety around trajectory-level prevention and checkable evidence, with harnesses blocking dangerous actions and proving that required actions occurred. →read the paper ToolHazard: Scaling Adversarial Environments for Security Evaluation and Alignment of LLM-Based Agents
Synthesizes stateful tool environments and environment-specific indirect prompt injections, then reuses the generated attacks as security-alignment data. →read the paper Tripwire: Triggering Aligned Refusal via Statistically Certified Safety Neurons
Identifies safety-specific neurons statistically and clamps them only on harmful inputs, reducing jailbreak success without continuously perturbing benign requests. →read the paper
| World models stop hiding state and dynamics inside pixels | Twin: Playing an Unknown Game with a Test-Time Digital Twin
Has a coding agent construct and continually repair an executable world model at test time, requiring it to reproduce observed transitions before relying on it. →read the paper Learning How the World Evolves: Extrapolative Video World Models via Latent Dynamics Reasoning
Encodes kinematic integration explicitly and learns only higher-order residual dynamics, improving extrapolation beyond trajectories represented during training. →read the paper Marionette: Predicting World States, Rendering Geometry, Painting Appearance
Separates a compact 3D world state, a fixed geometric renderer, and a diffusion appearance model so structural errors can be repaired directly in state. →read the paper RynnValue: Scaling Robotic Value Foundation Models with Temporal Distance
Turns temporal distance in large-scale robot video into a general value signal, converting weak timestamp supervision into dense rewards for robotic control. →read the paper
| Computation becomes selective, parallel, and structurally cheaper | Claim-Level Reliability Assessment for Efficient Test-Time Reasoning
Verifies decision-critical claims instead of sampling more complete solutions, using semantic falsification to locate decisive errors with fewer tokens. →read the paper Second Thought: Reasoning in Parallel as LLM Agents Act and Observe
Fills the idle time between agent actions and observations with parallel reasoning branches, reducing sequential decoding along the task’s critical path. →read the paper Multimodal Model Diffing for Feature Discovery and Control
Aligns and compares sparse autoencoders before and after multimodal training to isolate newly learned features that can be causally removed or steered. →read the paper Scaling Domain Data Repetition in LLM Pretraining
Shows that optimal repetition of scarce domain data increases mildly with model size and tracks domain validation loss better than the amount of unique data. →read the paper Dion3: Full-Stack Orthogonal Updates
Reworks orthogonalization, kernels, and distributed communication to cut optimizer-step time by as much as sixfold without worsening training loss. →read the paper
|
| |
|
| | | | | | |
|
