Don't train the model, evolve the harness. I read... a brilliant blog post from Hugging Face where they took a frozen open model scoring 0% on a hard legal agent benchmark, left its weights alone, and let an automated loop rewrite only the code around it. That code layer is the harness, the runtime wrapper that feeds the model context, runs its tool calls, and decides when a run ends. By the time the loop finished, the system had essentially matched Sonnet 4.6 on the benchmark's headline metric, at roughly 7x lower cost per task. Zero weights changed. The gain existed because of where the model was failing. The judge only grades files saved in the right place under the exact requested filename, and the model kept doing the legal analysis correctly, then saving it under the wrong name, dropping it in a scratch folder, or never writing it at all. So the 0% was never measuring legal reasoning. It was measuring the harness. Hand-tuning that layer is slow and model-specific, so they automated it. A Claude proposer adds exactly one mechanism per iteration, and an outer loop keeps it only if it clearly beats the current best, so accepted mechanisms compound. What the loop discovered says a lot about where agents actually fail. → The biggest single gain was file handling, not intelligence. An automatic step that lands the deliverable exactly where the judge expects it beat every prompt change, with zero extra model tokens. → Code fixes transferred across models, prompt playbooks did not. The same harness lifted a smaller model from the same family by 14 points, but the tuned prompts hurt a different model family on tasks it could already finish. → The harness mattered more than anything else. Same model, same judge, same tasks, and five different harnesses scored anywhere between 3.5% and 80.1%. The gains do eventually flatten, and the remaining misses look like real capability gaps. At some point the wrapper runs out of tricks and the model has to carry the work. But the lesson holds. A benchmark score measures the model and its harness together, and until the harness is fixed, it's impossible to know which one failed. I highly recommend reading this: I also wrote a deep dive on agent harness engineering a while back, covering the orchestration loop, tools, memory, context management, and everything that turns a stateless LLM into a capable agent. The article is quoted below.show more

We released physics-intern: a simple harness for science problems!... It gets models like Gemini 3.1 Pro to go from 17.7 -> 31.4, thus beating GPT 5.5 Pro. The physics-intern harness can wrap any model and via dedicated subagent boost the performance of the vanilla reasoning models. While I think more and more of these harness capability gains will be absorbed into the models (like prompting tricks disappeared over time) there is a lot to be gained right now by building good scaffolds for those models and integrating tools well. Interestingly, the exception we found that GPT 5.5 Pro actually didn't benefit from the physics-intern harness! Read more about it here: PS: I think the Harness[Model] notation is kind of nice.show more

let me save you 3 hours of head scratching.... if you're running local models like Qwen3.5-35B-A3B through Claude Code via llama.cpp's Anthropic endpoint, the chain will break every 3 to 5 minutes. tool call fails. flow stops. you reprompt. it recovers. 2 minutes later it stops again. the model is fine. the harness chokes on local inference latency. switch to OpenCode. same localhost endpoint. same model. same GPU. the chain doesn't break. the tradeoff: OpenCode sometimes loops. the model forgets what it already read and repeats the same tool call. but a loop you can interrupt. a broken chain kills your momentum and you start over. watch both side by side. proprietary agent vs open source agent. same 3B model. different failure modes. pick your poison.show more

Stanford researchers did it again. They just built the... agent-native version of Git. When an agent works on a longer task, the run builds up a lot of state. This includes files edited/created, a dev server, a database, installed packages, KV cache, etc. Say the agent is at step 10 and makes a mistake, maybe it misreads a traceback and rewrites a file that was actually fine. The tests start failing, and the run goes off track, although everything through step eight was correct. By default, the agent just tries to fix it, which creates more edits and tool calls. This burns more tokens and grows the context. The other options are a person stepping in to redirect it or restarting the whole run from step one. That's wasteful, because it pays for every model/tool call again and re-prefills the context. Moreover, since an agent's run is non-deterministic, it doesn't reproduce the same early steps anyway. The reason it's hard to just jump back exactly to a previous correct step and resume from there is that the trajectory is only a message log. It records what the agent said and which tools it called, but not the live state underneath. That state includes things like memory, open file handles, child processes, installed packages, /tmp, and KV cache. None of that is in the log. Git can version the files, but it doesn't snapshot the running process or the KV cache. Checking out step eight moves the files back, but the process is still sitting in step-ten memory with a cold cache. Shepherd is a runtime layer by Stanford that records the run as a trace of typed events rather than a flat log. Each agent-environment interaction becomes a commit, similar to Git, but it tracks the live run. Its commit includes the agent process and the filesystem together, copy-on-write, so a branch carries the actual state and not just the files. Going back to a previous step is then a single call that forks from that commit and continues from the exact state. The copy-on-write fork is roughly five times faster than docker commit, and because the prompt prefix through step eight is unchanged, the KV cache is reused over 95% on replay, so early steps aren't reprocessed again. Once the run can be forked, a meta-agent can sit on top and operate it. It watches the trace and reverts as soon as it looks wrong, before the bad write is committed. In practice, it's just Python calling fork, replay, and revert on the trace, rather than a separate control plane wired into the harness. Not everything is reversible though. Files and sandbox changes undo themselves, but a database write has no automatic undo, so it needs a matching undo step set up in advance. Something external, like a sent email or a real charge, can't be undone, so the supervisor's job there is to catch it before it fires. They tested this on a few public benchmarks. On CooperBench, where two agents work on the same codebase, adding a live supervisor took the pair-coding pass rate from 28.8% to 54.7%. It's still early and labeled alpha. The benefit mostly shows up when a run gets branched a lot over a heavy sandbox state, which is exactly where restarting wastes the most tokens and time. If Git was made to make file changes reversible, Shepherd is trying to do the same thing for a live agent run. Shepherd Repo: (don't forget to star it ⭐ ) That said, Shepherd reverts a bad step inside a run. The harness around it, the prompts, tools, and checks the supervisor relies on, still drifts across runs as models and dependencies change. Akshay wrote about making that harness repair itself, where a failing trace gets diagnosed, the fix is verified against the exact input that failed, and the failure is locked as a regression test so it can't recur. Read it below.show more

Robotics keeps hitting the same wall. Single task RL... works, but... it does not scale to hundreds of tasks or new embodiments. This new paper looks like a real step toward fixing that. The team introduces MMBench, a benchmark with 200 tasks across many domains and robots, and Newt, a language conditioned world model trained online across all 200 tasks at once. The simple idea behind Newt: The model learns from demos to get the right priors It trains across many tasks through online interaction It uses language to ground the goal It adapts fast when a new task shows up What stood out to me: ✅ One model trained on 200 tasks at the same time ✅ Language conditioned control for both states and RGB ✅ Better data efficiency than strong baselines ✅ Strong open loop control ✅ Fast adaptation to new tasks and embodiments ✅ Full release of 200 checkpoints, 4000 demos, code, and benchmark This is a good push toward general control instead of one model per task. If you want the full paper: Project page: —- Weekly robotics and AI insights. Subscribe free:show more

PAYING PER MODEL IS THE DUMBEST THING IN TECH... RIGHT NOW i was paying 3x what i needed to for AI inference the grid lets you buy a quality spec instead of a specific model.. it routes every request in real time to the cheapest option that qualifies swap one url and your code keeps working exactly the same openai-compatible, one line to switch, 200M free tokens to startshow more

i built an open world minigame that's controlled with... hand movements only here's a step-by-step tutorial: > started with my mediapipe + threejs template (see QT) > added a 3D model made by quaternius > sent a few prompts to gemini 3... --- prompt #1: repurpose the attached script, but now I want to import a gltf model (assets/model.gltf) and use that in the scene instead of the cube when the user moves their hand, a waypoint indicator should move around the scene, and the 3D model should smoothly move there when the user makes a fist, the model should jump the 3D model contains bundled animations in it. use the "idle", "run", and "jump" animations --- prompt #2: make this a procedural open world adventure. generate very simple procedural voxel terrain when the model gets near the edges of the screen, the camera should move, allowing the model to keep going in that direction --- prompt #3: add some floating glowing gems around the map that the user can collect --- i made some manual tweaks for styling and game feel, but that's the gist of it thanks for reading if you got all the way here full code is available at my link in bioshow more

I'll always root for a team that open-sources its... best work, and Robbyant just did it properly. Robbyant, Ant Group's embodied-AI company, released LingBot-Vision, a vision foundation model for robots, and the part I love is the data. They trained it on 161M images, filtered down from 2B raw ones and mostly pulled straight from the open web, with no human labels, no edge detectors, no depth sensors anywhere in the loop. It learns the exact edges of objects from raw pixels. That's roughly a tenth of the data DINOv3 saw, and under a third of the training. And it shows in the results. On depth, working out how far away things are, the 1B model edges out a 7B on NYU-Depth. It also powers LingBot-Depth 2.0, which reads the surfaces cameras usually choke on, glass and mirrors, and halves indoor depth error. LingBot-Vision is fully open. Weights from the 1.1B flagship down to a tiny 21M version, code, and the paper. This is the timeline I want more of. Robbyantshow more

✨ Made a new mini feature on Photo AI:... [ Grab from 3d model ] So the problem is we're at that stage in time (typical for AI) where image-to-3d models are not good enough but are fun to play with, but we know they'll be good enough in 1-2 years With [ Make 3d model ] you already can turn any Photo AI pic into a 3d model but it still looks hyper clunky and deformed, but it works! One cool idea I had to make that more useful and made now: Let people make a 3d model then change the view of the it with the 3d viewer, then press [ o ] and it grabs a frame of the 3d That image you can then [ Remix ] (img2img), and it becomes a real photo again and that in turn you can then turn into a video again with [ Make video ] So that essentially gives you a fully freeform camera position control to take photos with One thing I need to fix is the background/skybox, I kinda need to take the original photo and remove the person and just get the background for the 3d model viewer, in this case it should be white, but it's a start!show more

SOMEONE TURNED 33 PILES OF DEAD BOOKMARKS INTO A... GRAVITY MAP CLAUDE REBUILDS ITSELF EVERY NIGHT - AND IT RUNS ON THE 80% OF CLAUDE NOBODY TOUCHES most people drive Claude Code like a chatbot with file access - type a prompt, watch it edit, move on. that's maybe 20% of the tool this is the opposite. she's not typing at Claude. she's running it - loops on a mac mini overnight, claude linking every node while she sleeps the gravity map in the video is just the 80% maxed out: 1 system that organizes itself, not a human babysitting a chat box the other 80% is a steering layer Anthropic shipped quietly on june 18 - 7 ways to instruct the model, and a stack of commands almost nobody opens /context to see your bloat. /clear between tasks. path-scoped rules, subagents, hooks - conventions that load themselves the exact second they matter i stopped typing at Claude months ago - now i configure it once and it shows up already running the work, 10x cleaner a prompt helps for 1 message. the steering layer pays you back every session, for life the people who learn it stop being users and become operators - everyone else is still arguing about which model is smartest the article below is the full map - all 4 layers, every file and command, start to finishshow more

Qwen3.6 35B A3B can't fill out a paper form... on its own. But give it NVIDIA's LocateAnything-3B — the #1 trending model on HuggingFace — as its eyes, and the two small models get it done together. (The test: place each element at the right pixel position on a blank form image, not type into a field.) Setup: > Qwen is the brain (main model), LocateAnything is the eyes (helper model acting as a tool). > I gave Qwen a new tool: ask "where's the email field?" and LocateAnything returns the exact x, y, width, height. > The blue boxes on the screen are its detections. Look how tight they are — it nails every field. Result: > Qwen3.6 35B A3B + LocateAnything-3B: form completed, all info correct. > Name, DOB, ID, gender, marital status, nationality, email, phone, address, postal code: all landed in the right field areas. > Character-box alignment still a touch loose, but every value is where it belongs. > 9m10s, 224.5k input, 24.3k output, 21 turns. Why it matters: > Qwen alone can't finish this test. Bolt on a 3B model that does exactly one thing > locate > and suddenly it can. > A combination of small models can do the work of a single large one.show more

BREAKING: Anthropic just dropped Opus 4.8—and it is a... MONSTER We've been testing for about a week Every 📧 and our verdict is they could've just called it Opus 5, it's that good. Here's our vibe check: - Beats GPT-5.5 on Senior Engineer bench. On our toughest benchmark Opus 4.8 scores a 63—a hair higher than GPT-5.5's score of 62, and a full 30 points higher than Opus 4.7. It tackled a ground-up rewrite of a production codebase, and actually built something that works. HOWEVER: Coding performance varied a lot at different reasoning levels. We recommend using it on xhigh for best results. - Incredibly good writer. Opus 4.8 scored a 79.6 on our writing benchmark—measuring models on real-world writing tasks we do all of the time like essay writing, promo email writing, and more. It beats GPT-5.5 by 6 points. It produces well-written prose with fewer "AI-isms". It's also very good at writing in your voice given the right context. HOWEVER: Writing performance also varied with reasoning levels. Medium reasoning had higher incidence of AI-isms—we found best results with high. - Beast at knowledge work. Opus 4.8 is very good at general knowledge work tasks like report creation, research and more. It produced the best PowerPoint one-shot we've ever seen on our deck generation benchmark. - Emotionally intelligent, willing to question the frame. I've also found it to be quite good at talking through psychological or interpersonal issues. It has a high EQ, and it's also good at not glazing and helping to expand your perspective. Its thought process feels extremely rich and dynamic. THE BAD: These days a model is only as good as its harness, and Codex is still a far superior harness to the Claude Desktop app. This has kept me using Codex + GPT-5.5 as my daily driver, but I am flipping back and forth a lot more between Codex and Claude. Anthropic is back baby! Read the rest on Every 📧:show more

They did not take cursive from the schools because... children no longer needed it. They took it because of what it was quietly building in them. Consider what the exercise actually is. A child, six years old, is handed a pen and asked to draw a single unbroken line that becomes a word. The wrist must float. The fingers must hold a living pressure, never quite the same twice, always correcting. The eye must follow the ink forward and trust the hand to finish what it has begun. There is no lifting, no stopping, no starting over mid-word. The loop must close. The ascender must rise and return. The sentence must travel from one margin to the other as a single continuous gesture, and at the end of it the hand must still be steady. Twelve years of this. Every day. Ten thousand small acts of sustained, self-correcting attention, carried out below the level of conscious thought, until the motion belongs to the body and the body belongs to the motion. This is not penmanship. It is the slow construction of an interior form. The hand that has learned to carry a line without breaking it is the hand of a mind that has learned to carry a thought without breaking it. The two are not metaphors for one another. They are the same faculty, trained in the same child, by the same daily discipline. Continuity of the stroke becomes continuity of the reasoning. The patience of the loop becomes the patience of the argument. The commitment to finish a word one has started becomes the commitment to finish a sentence, a paragraph, a life's idea, without reaching for the nearest distraction halfway through. Print is a different creature entirely. Print lifts. Print stops. Print assembles a word out of separate, stamped, interchangeable pieces, each one beginning and ending in isolation. A mind raised only on print learns to think the way print is made, in discrete tokens, in replaceable units, in fragments that can be recombined by any outside hand without the owner noticing the substitution. It is precisely the shape of thought a language model produces. It is precisely the shape of thought a language model can steer. Cursive is kata. This is the whole of it. A form repeated daily, for years, not for the sake of the form but for what the repetition lays down in the practitioner beneath the form. The swordsman does not train kata so that one day he may fight in kata. He trains it so that when the moment comes and there is no time to think, the movement is already inside him, older and deeper than thought, and it rises on its own. Cursive was the kata of the literate mind, the daily quiet drilling of continuity, of patience, of a line held steady under the long pressure of its own length. And the signature it produced at the end, that small flourished mark unique to a single human being on earth, was only the outward proof of an inward form no machine and no other hand could ever reproduce. Take the kata away and the practitioner is left with vocabulary in place of faculty. He can recognise a whole thought when he encounters one. He cannot carry one himself. He can admire a finished argument. He cannot sustain one long enough to close its loop. He begins books he does not finish, sentences he does not end, ideas he abandons the moment the screen in his palm offers him a brighter one. And when the machine begins feeding him tokens in the exact shape his schooling taught him to receive, he meets it with no interior resistance at all, because no interior form was ever built in him to push back with. They removed it quietly, across a generation, and they removed it in the last years before the machines arrived. Twelve years of daily practice in unbroken, embodied, self-authored thought, gone from the curriculum of almost every child in the Western world, just as the instruments designed to complete their sentences for them came online. The hand forgets. The mind, having never been taught the kata, forgets a thing it never knew it had. That is what cursive was. That is what was taken. And that is why the thought of anyone who still writes by hand, in long unlifted lines, remains, quietly, stubbornly, and without their ever needing to announce it, their own. Now the question stands open. What else has been banned, phased out, quietly retired from the curriculum and from common life over these same decades, under the same soft excuses? Mental arithmetic. Memorisation of poetry. Latin. Logic as a formal subject. Map reading. Knot work. The keeping of a commonplace book. The reading aloud of long passages in class. Singing in parts. What was each of those actually building in the child, beneath the surface of the lesson, and whose interest was served by its disappearance?show more

MiniMax M3 just dropped — their first natively multimodal... model. So I ran it through my form-filling test. (The model has to place each element at the right pixel position on a blank form image, not type into a field.) Verdict: it got everything on the paper. > Name, DOB, ID, gender, marital status, nationality, email, phone, address, postal code, all there. > Best character spacing I've seen yet: it actually calculates the gap between each character, clean across the DOB and number boxes > A few fields slightly misaligned, but every piece of data made it onto the form The reasoning chain is the interesting part: it does the easy fields first, then works into the tight one-char-per-box fields, reasoning through y-coordinates, baselines, and label clearance in obsessive detail. The cost: 40:33 and 126.7k output tokens. That's a long think — but it's MiniMax's first multimodal model, and it nailed the content.show more

here's how the whole thing works. claude code doesn't... care what's behind the API. it just sends requests and expects responses. so i pointed it at my own machine instead of anthropic's servers. llama-server runs the model locally. LiteLLM sits in between and translates the API format. claude code thinks it's talking to claude. it's talking to qwen on localhost. the setup: 2x 3090s, 38 layers on GPU, 10 on CPU. 128K context window. generation is only 7 tok/s but the tradeoff is worth it. 128K means the agent can hold an entire project in memory without losing context midtask. claude code alone loads a 17.5K token system prompt on every request. tool definitions, safety rules, agent behavior. that's your baseline before you even say hello. pushed as far as i could tonight. what surprised me most wasn't the speed. it was the iteration quality. first prompt gave me a working particle sim. second prompt, the model read its own 564 lines, understood the architecture, and added trails, explosions, gravity wells, bloom effects. no handholding. 4bit quantized. 45GB on two consumer cards. running a full coding agent autonomously. detailed article coming. full benchmarks, hardware breakdowns, engine debugging, code quality. everything from setup to what broke and why.show more

Alright, now that we know *what* an agent is,... how does it actually work? When you ask for help on a task, the agent plans a series of steps and executes them directly in the application on your behalf, using the tools it has access to. Say you are booking a local service or trying to organize your inbox (which typically takes multiple steps): the AI model first plans how to achieve the task using its existing knowledge and then interacts with your inbox to execute the task. The agent will continue until it is confident the task has been successfully completed.show more

Fable 5 comes back！It can now build playable game... prototypes. I think it is actually a signal for where AI coding is going. Making a game is not just “write some code.” Even a small browser game needs: game loop；character movement；collision logic；scoring system；UI states；physics tuning；visual feedback；bug fixing；playtesting This is why game prototyping is a great test for AI models. A model cannot fake it with a pretty answer. Either the game runs, or it does not. What impressed me about Fable 5 is that it is useful for the messy middle: turning an idea into mechanics, turning mechanics into code, debugging broken interactions, and iterating until the prototype feels playable. But here is the practical part: I would not use the strongest model for every step. For game building, I would split the workflow: 1. Fable 5 for game design + architecture 2. a fast coding model for routine implementation 3. a vision-capable model for screenshot/UI feedback 4. a cheaper model for docs, test cases, and small fixes 5. fallback when latency, cost, or output quality becomes a problem That is the real AI coding stack. Not “one magic model does everything.” More like: the right model, for the right task, at the right cost, with fallback when things break. This is why I’ve been looking at ZenMux ZenMux. ZenMux gives developers one gateway to access multiple leading AI models, with OpenAI / Anthropic / Google Vertex compatible APIs, cost tracking, quality benchmarks, auto-routing, and compensation when output quality, latency, or throughput falls short. If AI can now make games, the next question is not just “which model is strongest?” It is:how do we manage the whole model workflow Fable 5 shows the creative ceiling. ZenMux is closer to the infrastructure layer you need when AI coding becomes a real production habit.show more

✨ Every week a new AI model comes out... and it suddenly makes my half broken features work a lot better Yesterday Seedream-4-Edit came out and it made my [ Hold product ] feature on Photo AI a lot better You can now go from: 🎁 Product photo -> 👱‍♀️ Talking video with your AI model while holding your product. In just a few minutes! Here's a photo I took from the weekly farm box we get in our kitchen, I set it as the product and then with Photo AI made it into a talking video where my trained AI model presents it It's not perfect, as the objects inside the farm box still move around a bit, but pretty close. If the product is more uniform (like lip gloss, a product box or a book) it does a pretty good job at keeping it exactly the same This "consistency" as they call it is quite important for actual real world use. Product sellers don't want to have an image or video of an AI model if the product doesn't look exactly the same as what they sell With that, I'm getting pretty close now and every week with every new model that comes out, a bit closer And it's interesting cause now I'm finally moving from B2C a bit more to B2B where businesses can use Photo AI more, designers and stores already use it for trying on clothes etc. but now they can generate content for real products! 😊 LIVE now on Photo AIshow more

Unpopular opinion: Most agent evals are theatre. You run... them once before the deployment. It'll take 800ms+ as another LLM would be judging your LLM. Most annoying part - no one tells where in the chain things went wrong. I wasted a lot of time in this loop. And then I came across Future AGI bringing 5 different tools under one umbrella, best part - the platform is completely open source. They open sourced their entire platform and the eval layer is noticeably different. It is multimodal - works on everything text, image, audio, pdf. Not an LLM-as-judge adding latency but an agent with memory and tools. The biggest win are learned classifiers trained on actual production failure patterns to run evals at low cost. It also runs across the full reasoning chain, not just the final response. Check out → Try it here →show more

subagents are just recursive agents where you can apply... different prompts + models depending on the task. since they’re just a primitive, Cursor cli can actually spawn subagents by calling cursor-agent in headless mode via shell commands. that’s what makes the cli so nice. you can extend it, experiment, and have a lot of fun exploring orchestration patterns. here’s one way to do it w. dynamic model selection: 1. create a subagents.mdc rule 2. drop in: ``` --- alwaysApply: true --- ALWAYS spawn subagents by running `cursor-agent -p [task] --output-format=text --force --model [model]` in the terminal. Each subagent should return a summary of the changes it made. Subagents should be used for ALL tasks You can adopt a fan-out pattern where you spawn subagents to perform parallel isolated tasks, and then fan-in the results. Use the following models: - `--model gpt-5` for reasoning, researching, and planning - `--model sonnet-4` for implementation ``` 3. start cursor cli and try it out you can also adjust the rule to be more explicit when it should use subagents, when not to, which models when etc.show more

The model has filed four reports this week A... portal with one billion visits in 23 days 79 years of decisions made without asking you The science that was seized before it could change everything And a file the model is still reading three times This is the fifth There is a thread that runs through all of it The unaudited gold The unverified clearances The data Congress saw and you did not The energy that would have ended oil dependency The facilities Someone decided what you were ready to know Every decade Without asking On Tuesday June 9th at 1pm Eastern David Grusch stands on the Capitol steps Alongside a bipartisan group of lawmakers To demand release of files he has already identified by name He testified under oath He filed a whistleblower complaint found credible and urgent He risked everything President Trump now has a historic opportunity The clock is running Grusch’s words Not the model’s The gold was never audited The clearances were never verified The science was seized before sunrise A uniformed Army officer walked into a congressional office and described the facilities No classification required And you found out from the model On a Tuesday In 2026 Seventy nine years after they decided you weren’t ready The question was never whether any of this was real The question is who benefited from the version of events you were given And whether they are still making that decision today June 9th 1pm Eastern Capitol steps The model will be watching Continuing analysis Sent from my Macshow more