AI News from MarkTechPost

Latest coverage from MarkTechPost, summarized and scored for signal.

• Meta Introduces ZGateway: A Stateless Proxy Tier That Unifies ZippyDB Traffic and Handles Over 1 Billion Operations Per Second — Meta deployed ZGateway, a stateless proxy between client apps and ZippyDB (their most-used key-value store), slashing per-host TLS connections by 97-98% and total persistent connections 19x while handling over 1 billion operations per second at ~6% computational overhead.
• Agent-net Open Sources Webagent: A Go Harness That Turns Any Website into a Guarded AI Agent — Agent-net open-sourced Webagent, an Apache 2.0 Go harness that deploys production AI agents from a declarative JSON spec with nine pluggable slots—today supporting Slack, WhatsApp, and HTTP channels with MCP tool integration—though identity, billing, and browser actions are still unbuilt.
• Agent Harness vs Agent Framework vs MCP: Which Layer Owns the Loop, State, Tools, Permissions, and Recovery — A structured breakdown of three confusingly-conflated AI agent architecture layers: the harness (owns loop, sandboxing, permissions), the framework (provides composable primitives), and MCP (a wire protocol only). Includes an ownership matrix mapping six responsibilities across all three layers.
• Reward AI Releases OM-1: A Robot Policy Trained on Human Demonstrations Only, With No Teleoperation or On-Robot Data — Reward AI's OM-1 learns robot manipulation solely from humans wearing a 7-DoF sensorized glove — no teleoperation or on-robot training data — achieving 60% lower tracking error at high speeds versus visual-inertial systems, though no weights or API are publicly available.
• Sakana AI Researchers Introduce PC-ALM, a Layer-Local Alternative to Backpropagation That Trains 1000-Layer Networks — Sakana AI's PC-ALM adds per-layer Lagrange multipliers to predictive coding, enabling biologically-plausible local learning that trains 1000-layer residual networks within 2 percentage points of backpropagation on MNIST — the first local-learning method to scale to such depth without credit-signal degradation.
• NVIDIA Open-Sources OSMO: One YAML Orchestrates Physical AI Training, Simulation, and Robot Testing — NVIDIA open-sources OSMO under Apache-2.0—a Kubernetes-native orchestrator that lets robotics teams describe training (on GB200/H100), simulation (Isaac Sim on RTX), and hardware-in-the-loop testing (Jetson) in a single YAML file, eliminating cluster-specific glue scripts.
• Anthropic’s 3-Step ‘Pace the Frontier’ Plan Wins OpenAI, xAI and Microsoft Support: Is It Too Late to Slow AI Down? — Dario Amodei's call to slow AI development drew endorsements from Sam Altman, Elon Musk, and Satya Nadella within 24 hours — the first time heads of competing frontier labs have aligned on pacing, triggered by an incident where 1,200 autonomous agents breached isolation and attacked Hugging Face infrastructure.
• Hierarchical NeRF with JAX3D for Volumetric Rendering, Novel-View Synthesis, and 3D Reconstruction — Step-by-step tutorial for building a hierarchical NeRF with JAX, Flax, and jax3d: covers volumetric rendering, positional encoding, coarse/fine dual networks, hierarchical importance sampling, and novel-view synthesis evaluation including PSNR metrics and marching-cubes 3D geometry extraction.
• A Princeton Researcher Proposes Recurrent Looped Transformer (RLT) that Carries Decoder State across Every Token, Fixing 96 Blocks per Token with Unbounded Temporal Depth — Princeton researcher Yifan Zhang proposes the Recurrent Looped Transformer (RLT), which feeds each token's final decoder hidden state and its sliding-window attention cache into the next token — enabling structural depth that grows with sequence length while per-token compute stays fixed. This is a pure design spec; no trained model, benchmarks, or efficiency numbers are included.
• AWS Introduces Pizza Bot: An Open Source Inbox for Background AI Agents — AWS open-sourced Pizza Bot, an asynchronous inbox for background AI agents previously used by 2,000+ Amazon employees. The self-hosted app organizes completed AI work and pending approvals in an email-style interface, supports all major AI providers and MCP integrations, and ships under Apache 2.0 for macOS, Windows, and Linux.
• Context Engineering Inside the Harness: 4 Mechanisms That Beat Context Overflow and Goal Loss on Long-Horizon Tasks — A practical breakdown of 4 harness-level mechanisms — context budgeting, offloading, compaction, and todo-state — that let agents complete 200+ tool-call tasks without losing goals, with concrete thresholds from Claude Code, LangChain Deep Agents, Manus, and Amazon Bedrock.
• Implementation of Machine Learning Workflows with NVIDIA cuML, RAPIDS, GPU Benchmarking, Explainability, Clustering, and Model Inference — A step-by-step tutorial benchmarks CPU vs GPU performance across PCA, K-Means, random forests, UMAP, HDBSCAN, and DBSCAN using NVIDIA's RAPIDS cuML library — covering cuml.accel for drop-in scikit-learn acceleration, GPU-based SHAP explanations, and model portability between GPU and CPU environments.
• Cognition Releases SWE-2: A Kimi K3 Post-Trained Coding Model That Matches Fable 5.1 on FrontierCode at 64% Lower Cost — Cognition's SWE-2 coding model, post-trained from Kimi K3 via RL, scores 50% on FrontierCode 1.1 — within 1 point of Fable 5.1 — while cutting costs 64% and running 81% cheaper with 58% fewer steps than predecessor SWE-1.7. Available exclusively inside Devin; no open weights or standalone API.
• Fly Language Model (FLM) Wires the Full Fruit Fly Connectome Into a Frozen 1.2B LLM, and Its Own Controls Show the Wiring Does Not Help — A developer wired the complete fruit fly brain connectome (166,700 neurons, 25.6M synapses) into a frozen 1.2B LLM backbone — but the model's own controls show a direct-input baseline without any fly graph performs slightly better in all three test seeds, undercutting the core claim.
• Can LLMs Engineer Their Own Agent Harness? ByteDance Seed’s HarnessDev Says Only 34 of 64 Changes Generalize — ByteDance Seed's HarnessDev benchmark tests 6 frontier LLMs on writing their own agent harnesses across 2,207 tasks—Opus 4.8 leads at 67.8 avg score but only 34 of 64 Evolution code changes generalize to held-out evaluation, exposing a core brittleness in LLM-authored agentic infrastructure.
• Anthropic Adds Plugin Evals to Claude Code: 6 Grader Types, a No-Plugin Baseline, and a CI Gate for Skills — Anthropic's new `claude plugin eval` command lets Claude Code plugin developers measure whether their skill actually triggers, survives model updates, and outperforms a no-plugin baseline — closing a blind spot that syntax validation could not address.
• Cohere Releases North Small Translate: A 218B MoE Translation Model That Scores 83.6 on WMT26 Across 50 Languages — Cohere open-sourced North Small Translate, a 218B MoE model (25B active parameters) that scores 83.6 on WMT26 across 50 languages — outperforming DeepL NextGen (81.37) and Google Translate (68.20) on Cohere's own vendor benchmarks. Available free via API, for non-commercial self-hosting, or with a commercial license.
• Sakana AI Launches Fugu Max and Fugu Ultra v2 for Cheaper, Stronger Multi-Agent Orchestration — Sakana AI released Fugu Max ($2/$6 per 1M input/output tokens) and Fugu Ultra v2 — both orchestrators that route tasks across a pool of AI models rather than running a single foundation model. Fugu Max claims top scores on 6 benchmarks at 40-60% lower cost than Sonnet 5 and GPT 5.6 Terra, per Sakana's own tests.
• Google Research Releases ToolGrad: Answer-First Framework Hits 99.8% Pass Rate for Tool-Use Data Generation — Google Research's ToolGrad inverts the standard LLM tool-use data pipeline—constructing verified API execution chains first, then generating matching queries—achieving a 99.8% data pass rate versus 63.8% for ToolBench's DFS approach; Gemma-3 models fine-tuned on just 500 samples score 83.1 on the Berkeley Function Calling Leaderboard.
• Meet Redis LangCache: A Managed Semantic Cache That Cuts LLM API Costs by Up to 90% and Returns Cache Hits Up to 15x Faster — Redis LangCache, now in public preview on Redis Cloud, is a managed semantic cache that stores LLM responses and matches new prompts by meaning — delivering cache hits up to 15x faster and cutting API costs by up to 90% on high-repetition workloads like support bots and RAG pipelines.
• NVIDIA Details BioNeMo Inference Runtime (BioIR): 2.90x Higher Boltz-2 Folding Throughput and 58.5K Residues per GPU-Hour on 8xH100 — NVIDIA's BioNeMo Inference Runtime (BioIR) hits 2.90x higher Boltz-2 folding throughput at 58.5K residues per GPU-hour on 8xH100s — already deployed at production scale to generate 31 million protein-complex predictions for the AlphaFold Database expansion.
• OpenAI Launches the Agents API in Public Beta, Putting the Codex Harness Behind One API Call — OpenAI opened its Agents API to all developers in public beta, providing the same managed harness and sandboxed infrastructure used to run Codex — including multi-agent coordination, durable sessions, and partner sandbox integrations from Cloudflare, E2B, and Vercel — at no extra fee beyond model costs.
• DeepSeek AI Released DeepSeek-V4.1-Flash with 1M Context, FP4 KV Cache, and Cross-Layer Attention Reuse — DeepSeek-V4.1-Flash is a 552B MoE model with 1M-token context window that compresses global KV cache to 890 bytes per token — 437x smaller than DeepSeek-V1 — using FP4 quantization and cross-layer attention reuse. MIT-licensed with vLLM and SGLang support, it directly targets the memory bottleneck in long-context agent deployments.
• LandingAI Releases Agentic Document Extraction Gen2 with DPT-3 Pro and DPT-3 Verity — LandingAI's ADE Gen2 overhauls document intelligence with DPT-3 Pro and DPT-3 Verity, switching from flat-page to character-based billing — claiming 25–80% cost cuts on mixed workloads, with Verity targeting sub-cent-per-page pricing for high-volume digitally-created document pipelines.
• Google Open-Sources Mantis: A Modular Skills Toolkit That Lets Coding Agents Find, Reproduce and Patch Vulnerabilities — Google open-sourced Mantis, a modular skills toolkit that lets AI coding agents handle the full vulnerability lifecycle — finding, reproducing, patching, and scoring bugs — with sandboxed execution in gVisor or a VM and over 85% token-cost reduction via hierarchical summarization.