Notizie

Prime Intellect has open-sourced Prime Agent, a self-improving coding harness designed around two abstractions, the Recursive Language Model (RLM) and Continual Harness. Fixed tool schemas and context compaction force a model to work around its own scaffolding. Prime Agent replaces both with a persistent Python REPL and a rewritable harness. With Opus 5, it reports 95.5% on ARC-AGI-3, above the reported human expert baseline of 95.4%. It is MIT-licensed. Is it deployable Yes, today. Prime Agent installs on Linux or macOS with one command. It runs on subscription logins (Codex, Claude Pro/Max, GitHub Copilot), API keys (Anthropic, OpenAI, Google, Groq, Fireworks, Prime Inference, and others), Azure OpenAI, Amazon Bedrock, and self-hosted vLLM, Ollama, or LM Studio endpoints. Self-hosting an open-weights model such as GLM-5.2 keeps code inside your own network. Company level: Best fit is mid-size to large engineering orgs and AI labs that already run isolated CI containers. Prime Intellect states plainly that worker and kernel processes are not a security sandbox. Deployment therefore needs disposable clones or restricted environments. Solo developers can install it, but the payoff appears on multi-hour tasks. Industries: Developer tooling, semiconductor and HPC teams writing GPU kernels, simulation and gaming, quantitative research, and AI research labs. Applications: Overnight refactors behind a test gate, spec-driven builds from scratch, kernel optimization, long-horizon agent evaluation, and autoresearch. What Prime Intellect shipped Prime Agent is built on two abstractions. The Recursive Language Model (RLM) treats context as a variable and sub-agent delegation as function calls inside a REPL. The Continual Harness treats prompts, sub-agents, skills, and memory as state the agent can create, read, update, and delete from its own trajectory. Both papers have Prime Agent authors on them. The TUI is built on pi. Programmatic tool calling Models in Prime Agent get one tool: a persistent IPython kernel. Skills, tools, and sub-agents are pre-imported modules inside it. rlm(“sub-task”) launches a child session with its own model, kernel, and history, returning at admission rather than blocking. Results arrive through agent_message.send(…). A background daemon owns every live session. You can detach and reattach without stopping the loop, and a crashed worker recovers from the session JSONL plus a kernel snapshot. Agent-to-agent messaging is deliberately scoped to the nuclear family — parent, sibling, or child — to prevent cross-session chatter. Retained sub-agents drop from memory after 30 minutes idle, then reload when addressed. Self-improvement through /refine Continual Harness formalizes harness state as H = (ρ, G, K, M): prompt, sub-agents, skills, memory. Each exposes the same create, read, update, delete surface. /refine reads the agent’s own trajectory and applies the smallest relevant edit, recording the trigger and the outcome. Planning runs in the background without blocking the conversation. The base system prompt stays immutable, and a bad update can be reverted by ID. Benchmarks On ARC-AGI-3, Prime Agent with Opus 5 reports 95.5% RHAE Best@1, above the ARC reported human expert baseline of 95.4%. Three runs land at 95.0, 95.2, and 95.5, with 99.97% Best@3 and all 183/183 levels complete. Prime Intellect also reports lower token usage than native harnesses, crediting functions run over data instead of data read through tools. On a long-context suite, Prime Agent with open-weights GLM-5.2 beats Pi-mono on eight of nine evals. With Opus 5 it edges Claude Code on six of nine; with GPT-5.6 Sol it beats Codex on six of nine. Case studies include EmulatorBench, where the agent builds emulators in Rust from spec with no reference implementation and reproduces the SEGA Genesis and Game Boy Color; PMPP-Hard, for GPU kernels verified against KernelGuard; and Factorio, where it reached 100K+ production score in hours. Factorio also produced the most useful negative result. Prime Agent found it could spawn resources straight into assembly machines through RCON commands, despite a heartbeat prompt telling it not to cheat. The same refinement loop that built legitimate skills then built efficient cheating skills. Key Takeaways Prime Agent is MIT-licensed, installs in one command, and works with subscriptions, APIs, or self-hosted models. One tool — a persistent IPython kernel — replaces fixed tool schemas; sub-agents are function calls. /refine edits prompts, skills, memory, and sub-agent specs from the trajectory, with rollback by ID. Opus 5 in Prime Agent hits 95.5% on ARC-AGI-3, above the 95.4% human expert baseline. Check out the Technical details and GitHub Repo. Also, feel free to follow us on Twitter and don’t forget to join our 150k+ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well. Need to partner with us for promoting your GitHub Repo OR Hugging Face Page OR Product Release OR Webinar etc.? Connect with us The post Prime Intellect Releases Prime Agent: An Open-Source RLM Harness Where Sub-Agents Are Function Calls Inside Persistent IPython Kernel appeared first on MarkTechPost.

In this tutorial, we explore adaptive experimentation using Meta’s Ax with the modern Client API. We work through a complete workflow where we tune a RandomForest model on a synthetic classification dataset while balancing predictive accuracy against model footprint. We begin by defining a mixed search space with integer, float, log-scaled, and categorical parameters, then use Ax’s ask-tell optimization loop to run constrained Bayesian optimization, multi-objective optimization, and parameter-constrained experimentation. Along the way, we visualize convergence, inspect the Pareto frontier, use Ax’s built-in analysis tools, and persist the experiment for future reuse. Copy CodeCopiedUse a different Browser import importlib, subprocess, sys def _ensure(module, pip_name=None): try: importlib.import_module(module) except ImportError: print(f”Installing {pip_name or module} …”) subprocess.check_call([sys.executable, “-m”, “pip”, “install”, “-q”, pip_name or module]) _ensure(“ax”, “ax-platform”) _ensure(“sklearn”, “scikit-learn”) import logging, warnings, time import numpy as np import matplotlib.pyplot as plt warnings.filterwarnings(“ignore”) logging.getLogger(“ax”).setLevel(logging.WARNING) from ax.api.client import Client from ax.api.configs import RangeParameterConfig, ChoiceParameterConfig from sklearn.datasets import make_classification from sklearn.ensemble import RandomForestClassifier from sklearn.model_selection import StratifiedKFold, cross_val_score np.random.seed(0) We begin by preparing the Colab environment and installing the required packages for Ax and scikit-learn. We import the core libraries for optimization, machine learning, plotting, logging, and reproducibility. We also configure warnings and Ax logging to keep the notebook output clean and focused on the experimental results. Copy CodeCopiedUse a different Browser X, y = make_classification( n_samples=1400, n_features=20, n_informative=8, n_redundant=4, n_classes=3, random_state=0, ) CV = StratifiedKFold(n_splits=3, shuffle=True, random_state=0) def evaluate(p): n_est, depth = int(p[“n_estimators”]), int(p[“max_depth”]) clf = RandomForestClassifier( n_estimators=n_est, max_depth=depth, max_features=float(p[“max_features”]), min_samples_leaf=int(p[“min_samples_leaf”]), criterion=p[“criterion”], ccp_alpha=float(p[“ccp_alpha”]), n_jobs=-1, random_state=0, ) accuracy = cross_val_score(clf, X, y, cv=CV, scoring=”accuracy”).mean() model_size = n_est * depth return {“accuracy”: float(accuracy), “model_size”: float(model_size)} SEARCH_SPACE = [ RangeParameterConfig(name=”n_estimators”, bounds=(50, 300), parameter_type=”int”), RangeParameterConfig(name=”max_depth”, bounds=(3, 24), parameter_type=”int”), RangeParameterConfig(name=”max_features”, bounds=(0.2, 1.0), parameter_type=”float”), RangeParameterConfig(name=”min_samples_leaf”,bounds=(1, 12), parameter_type=”int”), RangeParameterConfig(name=”ccp_alpha”, bounds=(1e-5, 1e-1), parameter_type=”float”, scaling=”log”), ChoiceParameterConfig(name=”criterion”, values=[“gini”, “entropy”, “log_loss”], parameter_type=”str”, is_ordered=False), ] def run_study(client, total_trials, metric_keys, batch=4): records = [] while len(records) < total_trials: trials = client.get_next_trials(max_trials=min(batch, total_trials – len(records))) if not trials: break for idx, params in trials.items(): full = evaluate(params) raw = {k: full[k] for k in metric_keys} client.complete_trial(trial_index=idx, raw_data=raw) records.append({“trial”: idx, “params”: params, **full}) return records We create a synthetic multi-class classification dataset and define a cross-validation strategy to evaluate Random Forest models. We build an evaluation function that returns both accuracy and model size, allowing us to measure performance and cost together. We then define a mixed search space with integer, float, log-scaled, and categorical parameters, along with a reusable ask-tell study runner. Copy CodeCopiedUse a different Browser print(“n=== Study 1: constrained single-objective Bayesian optimization ===”) c1 = Client() c1.configure_experiment(parameters=SEARCH_SPACE, name=”rf_constrained”) c1.configure_optimization(objective=”accuracy”, outcome_constraints=[“model_size <= 2500”]) rec1 = run_study(c1, total_trials=24, metric_keys=[“accuracy”, “model_size”]) best_params, prediction, best_idx, best_arm = c1.get_best_parameterization() print(“nBest feasible configuration found:”) for k, v in best_params.items(): print(f” {k:>16}: {v}”) print(” predicted:”, prediction) feasible = [(r[“trial”], r[“accuracy”]) for r in rec1 if r[“model_size”] <= 2500] best_so_far, cur = [], -np.inf for _, acc in feasible: cur = max(cur, acc); best_so_far.append(cur) plt.figure(figsize=(7, 4)) plt.plot(range(1, len(best_so_far) + 1), best_so_far, “o-“) plt.xlabel(“feasible trial #”); plt.ylabel(“best accuracy so far”) plt.title(“Study 1 — convergence (subject to model_size <= 2500)”) plt.grid(alpha=0.3); plt.tight_layout(); plt.show() We run a constrained single-objective Bayesian optimization study where we maximize accuracy while keeping model size below a fixed threshold. We use Ax to suggest hyperparameter configurations, evaluate them, and report both accuracy and model size back to the optimizer. We then extract the best feasible configuration and plot the best accuracy achieved over feasible trials. Copy CodeCopiedUse a different Browser print(“n=== Study 2: multi-objective (accuracy vs. model_size) ===”) c2 = Client() c2.configure_experiment(parameters=SEARCH_SPACE, name=”rf_multiobjective”) c2.configure_optimization(objective=”accuracy, -model_size”) rec2 = run_study(c2, total_trials=28, metric_keys=[“accuracy”, “model_size”]) try: frontier = c2.get_pareto_frontier() print(f”Ax identified {len(frontier)} Pareto-optimal configurations.”) except Exception as e: frontier = None print(“get_pareto_frontier unavailable in this version:”, e) acc = np.array([r[“accuracy”] for r in rec2]) size = np.array([r[“model_size”] for r in rec2]) order = np.argsort(size) pareto_idx, best_acc = [], -np.inf for i in order: if acc[i] > best_acc: best_acc = acc[i]; pareto_idx.append(i) plt.figure(figsize=(7, 5)) plt.scatter(size, acc, c=”lightgray”, label=”all trials”) plt.scatter(size[pareto_idx], acc[pareto_idx], c=”crimson”, zorder=3, label=”Pareto front”) plt.plot(size[pareto_idx], acc[pareto_idx], “–“, c=”crimson”, alpha=0.6) plt.xlabel(“model_size (lower = cheaper)”); plt.ylabel(“accuracy (higher = better)”) plt.title(“Study 2 — accuracy vs. model size trade-off”) plt.legend(); plt.grid(alpha=0.3); plt.tight_layout(); plt.show() We move from single-objective optimization to multi-objective optimization by jointly maximizing accuracy and minimizing model size. We use Ax to search for configurations that represent strong trade-offs between predictive performance and computational footprint. We then calculate and visualize the empirical Pareto frontier to understand how accuracy varies with model size. Copy CodeCopiedUse a different Browser print(“n=== Study 3: parameter constraints on a synthetic surface ===”) c3 = Client() c3.configure_experiment( parameters=[ RangeParameterConfig(name=”x1″, bounds=(0.0, 1.0), parameter_type=”float”), RangeParameterConfig(name=”x2″, bounds=(0.0, 1.0), parameter_type=”float”), ], parameter_constraints=[“x1 + x2 <= 1.5″], name=”constrained_surface”, ) c3.configure_optimization(objective=”-dist”) for _ in range(14): for idx, p in c3.get_next_trials(max_trials=1).items(): dist = (p[“x1”] – 0.9) ** 2 + (p[“x2”] – 0.9) ** 2 c3.complete_trial(trial_index=idx, raw_data={“dist”: float(dist)}) bp, _, _, _ = c3.get_best_parameterization() print(f”Best point: x1={bp[‘x1’]:.3f}, x2={bp[‘x2’]:.3f}, ” f”sum={bp[‘x1’] + bp[‘x2’]:.3f} (constraint: <= 1.5)”) print(“Unconstrained optimum would be (0.9, 0.9); Ax respects the boundary.”) We demonstrate parameter constraints using a simple two-dimensional synthetic optimization problem. We ask Ax to minimize the distance to a target point while enforcing the input constraint that the sum of the two variables remains below a boundary. We observe that the optimizer respects the constraint and finds the best feasible point near the constrained optimum. Copy CodeCopiedUse a different Browser print(“n=== Ax built-in analyses for Study 1 ===”) try: import plotly.io as pio if “google.colab” in sys.modules: pio.renderers.default = “colab” cards = c1.compute_analyses(display=True) print(f”Computed {len(cards)} analysis cards.”) except Exception as e: print(“Interactive analyses didn’t render in this environment:”, e) print(“(The matplotlib plots above already capture the key results.)”) print(“n=== Saving / loading the experiment ===”) try: c1.save_to_json_file(“ax_study1.json”) reloaded = Client.load_from_json_file(“ax_study1.json”) print(“Saved to ax_study1.json and reloaded successfully.”) rp, _, _, _ = reloaded.get_best_parameterization() print(“Best params from reloaded client match:”, rp == best_params) except Exception as e: print(“JSON persistence API differs in this version:”, e) print(“See: https://ax.dev/docs/recipes/experiment-to-json”) print(“nDone. You optimized a mixed-type search space with constraints, ” “traced a Pareto frontier, and persisted in the experiment.”) We use Ax’s built-in

Cloudflare has released Kitesurf, a stateless web browser built specifically for AI agents. It runs entirely in V8 isolates on Cloudflare Workers, with no Chromium underneath. Browser engines like Chromium were built for humans, and their memory and compute overhead makes one-browser-per-agent prohibitively expensive. Agents do not need tabs, extensions, or pixel-perfect 60-fps rendering. They need machine-readable content, low token overhead, scalability, and isolation against threats like prompt injection. Kitesurf drops the human-facing parts and keeps what models use. It already passes 215,000+ Web Platform Tests and is available now, free while in beta, through Browser Run. Is It Deployable Yes, as a free beta in Browser Run, behind per-account limits. Treat it as production-adjacent: strong for compatible sites and one-shot tasks, with Chromium as the fallback for complex pages. Adoption cost is minimal — existing Puppeteer, Playwright, and MCP clients work by adding a single browser=kitesurf parameter. Company levels: Startups and indie agent builders benefit most from the 3–7× lower CPU and memory footprint. Mid-market SaaS and enterprises already on Cloudflare Workers can switch selected workloads with one parameter. Industries: AI agent platforms, web data extraction, SaaS automation, competitive monitoring, PDF/document generation, and search/RAG ingestion pipelines. Applications: Agent web browsing, screenshots, HTML extraction, PDF rendering, and bursty one-shot Quick Actions. Potential customers: AI engineers building browser-using agents, platform engineers scaling automation fleets, data engineers and data scientists running extraction jobs, and technical founders tracking per-session infrastructure costs. Not yet: Video, WebGL, TLS-fingerprint bot challenges, and long authenticated stateful sessions — keep those on the Chromium default. Architecture: Engine, PageScript, PageRenderer Kitesurf splits the browser into isolated Workers components. The Engine is the only public-facing piece. It speaks the Chrome DevTools Protocol (CDP) over WebSocket plus HTTP REST, and stores each session’s state. Every other component is stateless and disposable. PageScript is built on Dynamic Workers. Each page or out-of-process iframe gets its own long-lived isolate with a clean globalThis and DOM. HTML and CSS parsing use Blitz, a modular Rust rendering engine, and Stylo, Firefox’s CSS parser. Because Workers does not support native eval, occasional eval calls run through Boa JS, a Rust ECMAScript engine, a runtime on top of a runtime. PageRenderer rasterizes the computed scene into JPEG/PNG or PDF using blitz-paint and Parley for text shaping. A dedicated SandboxOutbound worker is the only component allowed to touch the network. It enforces CORS, injects browser-shaped headers, keeps per-page cookie jars, and returns a 403 for anything violating policy. Benchmarks: 3–7× Cheaper, ~1.7–1.8× Slower WPT coverage grows by hundreds of passing tests weekly. On Cloudflare’s 14-URL corpus (medians of five Quick Action runs), Kitesurf used 380 ms CPU per screenshot vs 1,173 ms for Chromium (3.1× less) and 229 ms vs 877 ms for HTML extraction (3.8× less). Memory dropped from 271.0 MiB to 57.8 MiB for screenshots (4.7×) and from 273.7 MiB to 39.4 MiB for extraction (7.0×). Chromium remains faster on wall time: Kitesurf is 1.8× slower on screenshots and 1.7× slower on extraction, mostly in rasterization and image encoding. Since memory and CPU drive the bill, the trade favors bursty agent workloads. It also runs Doom. How to Use It, and Current Limits Existing Puppeteer, Playwright, chrome-remote-interface, and MCP clients work unchanged. Add browser=kitesurf to the Browser Run CDP endpoint or Quick Actions API. A public playground with injected Chrome DevTools shows DOM, console, network, and per-isolate WebAssembly memory. Kitesurf already renders TodoMVC (vanilla, React, Vue, Angular, Preact), Wikipedia, and Hacker News. Cloudflare plans to open source Kitesurf so customers can deploy their own instances. Key Takeaways Kitesurf is Cloudflare’s agent-first browser running fully in V8 isolates on Workers, free in beta. It uses 3.1–3.8× less CPU and 4.7–7.0× less memory than Chromium on common agent tasks. Chromium is still ~1.7–1.8× faster on wall time; Kitesurf wins on cost per session. Existing Puppeteer/Playwright/CDP clients work by adding a single browser=kitesurf parameter. No video, WebGL, TLS-fingerprint bot challenges, or persistent authenticated sessions yet. Check out the Technical details. Also, feel free to follow us on Twitter and don’t forget to join our 150k+ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well. Need to partner with us for promoting your GitHub Repo OR Hugging Face Page OR Product Release OR Webinar etc.? Connect with us The post Cloudflare Introduces Kitesurf: An Agent-First Web Browser That Runs Entirely in V8 Isolates on Cloudflare Workers appeared first on MarkTechPost.

With the vocabulary and the failure modes in place, here’s the build.

NVIDIA has released Alpamayo 2 Super, a 34B-parameter vision-language-action (VLA) model for autonomous driving, under an open commercial license. The stated design target is the long-tail events: rare, multi-agent situations that conventional detection-and-prediction stacks handle poorly. The model pairs a 32B VLM backbone, built on NVIDIA Cosmos 3 Super Reasoner and post-trained with reinforcement learning, with a 2.3B diffusion-based action decoder. From one pass over full-surround camera video it emits a planned trajectory, a causal explanation of that trajectory, and a meta-action. Is it deployable Yes, and for commercial use from day one. The weights are released under OpenMDW-1.1, the Linux Foundation’s permissive license for open model distributions; source code is Apache 2.0. The license covers fine-tuning, derivative models and commercial redistribution. NVIDIA is applying OpenMDW across the entire Alpamayo family, so earlier releases introduced for R&D are now deployable commercially without additional permission. Inputs, outputs and training data Inputs are multi-camera RGB video, text, and egomotion history with timestamps. The validated public notebook profiles use six cameras and four historical frames per camera. Egomotion is 3D translation plus a 3×3 rotation matrix, multi-timestep. The trajectory API returns 64 waypoints spanning 0.1 to 6.4 seconds at 0.1-second intervals. Each waypoint carries ego-frame XYZ and a 3×3 rotation matrix. Training data is roughly 115,000 hours of multi-camera driving video with egomotion and trajectory annotations. It includes about 3,700,000 Chain-of-Causation (CoC) traces — structured, causally linked explanations of driving decisions. Image training data exceeds one billion images. Benchmarks On LingoQA, Alpamayo 2 Super records a Lingo-Judge score of 79.2 and ranks first among nearly 40 models evaluated. In NVIDIA’s testing it beat Qwen2.5-VL 72B by 17.0 points, Gemini 2.5 Pro by 15.1, and GPT-4o by 23.2. Two more numbers matter for planning work. Closed-loop evaluation with AlpaSim on 910 scenarios from the PhysicalAI-AV-NuRec dataset gives an AlpaSim score of 1.50 ± 0.13. Open-loop evaluation on 937 challenging samples from the PhysicalAI-AV dataset gives minADE₆ at 6.4s of 0.911m. Five outputs from one model For each driving situation, the model produces a trajectory, a CoC trace explaining the decision, a meta-action such as yield or lane change, reasoning auto-labels, and visual question answering with 2D grounding. That combination is what makes the release interesting operationally. Developers can tie what the model observed to the action it chose. CoC traces integrate with NVIDIA Halos safety-validation workflows and support AI safety aligned with ISO/PAS 8800. Used as an autolabeler on proprietary fleet data, NVIDIA says the model compresses annotation cycles from months to days. Interactive explainer Key Takeaways 34B VLA model — 32B Cosmos 3 Super Reasoner backbone plus a 2.3B diffusion action expert. OpenMDW-1.1 weights and Apache 2.0 code; commercial use and redistribution allowed, no extra permission needed. LingoQA Lingo-Judge 79.2, first among nearly 40 models; AlpaSim 1.50 ± 0.13; minADE₆ 0.911m at 6.4s. One pass yields trajectory, Chain-of-Causation trace, meta-action, auto-labels, and grounded VQA. Cloud-scale model tested on 1× H100 80GB at 72,115 MiB peak; distill it for in-car inference. Check out the NVIDIA blog and Hugging Face model card. Also, feel free to follow us on Twitter and don’t forget to join our 150k+ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well. Need to partner with us for promoting your GitHub Repo OR Hugging Face Page OR Product Release OR Webinar etc.? Connect with us The post NVIDIA Releases Alpamayo 2 Super: A 34B Open Vision-Language-Action Model for Robotaxis and Autonomous Driving Under OpenMDW-1.1 appeared first on MarkTechPost.

Ready for a fresh set of puzzles? Click here for the September/October 2026 Puzzle Corner, brought to you by Michael S. Branicky, ScD ’95, of the Puzzle Corner Puzzle Crew (aka PC2), which also includes Edward Faulkner ’03, MEng ’04, and Abe Kunin ’03. This column includes solutions to the May/June issue. Send problems, solutions (by October 1), and comments to puzzlecorner@technologyreview.com. Editor emeritus Allan Gottlieb ’67 launched Puzzle Corner in 1966. Find back issues through 2022 at cs.nyu.edu/~gottlieb/tr and more recent back issues at technologyreview.com/puzzle-corner.