Weekly Tech Digest: OpenAI's Codex Harness, AI Agents, Robotics & Math Breakthroughs

Big Tech Updates

OpenAI Open-Sources Codex Harness

OpenAI released the core framework behind its Codex coding agent, Harness, under Apache-2.0. Harness is an execution system handling task understanding, memory, tool calling, error handling, and human-in-the-loop approval. Benchmarks show tuning Harness's reasoning retention and context compression strategies significantly improves model scores while cutting token usage. Three components are provided: a CLI for automation pipelines, an SDK for programmatic agent control, and an app-server for deep product integration with state persistence, event streaming, and approval workflows. The framework moves beyond chat-style AI, letting developers embed agents into existing business UIs while retaining control over context, tool access, and risk approvals. Case studies include a tax-filing workflow compressing processing time, Cisco building internal natural-language app generation, and a logistics dashboard where operators issue commands that trigger internal tools with human confirmation before execution.

DeepSeek Harness Adds Multimodal Support

DeepSeek Harness updated to support native image requests via model adapters, though DeepSeek v4 vision model is not yet official; third-party vision models can be plugged in. Instructions (goal, plan) now accept images alongside text, and @-references extend to historical conversations. Underlying MCP and ACP protocols now persist image attachments across calls, and PTC Mode supports nested image forwarding. Fixes address large-image request failures and multi-turn payload accumulation errors. Additional changes: Claude Code and Codex become optional install packages; Codex gains non-interactive permission mode for unattended automation and multi-named-instance support for running multiple sub-agents; sub-agent feedback wakes parent tasks promptly; web_search adds concurrent queries; Windows PTY adds PowerShell; minimal mode defaults to persistent bash and file editor; SQLite backend improves read/write and fork performance with reduced storage (breaking backward compatibility). DeepSeek Harness trademark registered.

Cursor Launches Origin Code Hosting Platform

During a near-seven-hour GitHub outage (three hours core downtime), Cursor unveiled Origin, a Git platform built for the agent era. Origin mirrors GitHub basics (repo creation, PRs, permissions) with real-time two-way sync and one-click migration. Differentiators target AI collaboration: stacked PRs visualize dependent changes; merge queues auto-sort and detect conflicts with built-in AI resolution; review state exposed as structured API for agents; native MCP support; integrations with Vercel, Depot, Buildkite; GitHub Actions compatibility. Architecture targets high-frequency agent cloning, committing, merging — Cursor reports a significant share of PRs already fully automated by cloud agents. Traditional GitHub workflows center on human review cadence, ill-suited for second-scale agent iterations. Origin unifies repo, PR management, and agent runtime, eliminating cross-platform hops. Large-team migration unlikely immediately, but Origin offers a viable alternative for Cursor cloud-agent users.

Industry Insights

Alibaba AI: Cloud Profits vs. Qwen Losses

Alibaba's Q2 FY2026 (ended June 30) revenue ¥268.953B (+9% YoY), but profits dipped due to heavy AI capex and R&D. Reorganized into three segments: E-commerce Group, AI Cloud & Compute, AI Lab & Applications. AI Cloud & Compute revenue grew 45% YoY; AI product revenue 12th consecutive quarter of triple-digit growth; segment profit rose on high-margin business, custom chips, and operational efficiency. Conversely, AI Lab & Applications revenue limited, losses widened sharply from model R&D and Qwen inference costs. Overall AI segment still investment-heavy. T-Head chips commercialized across industries; Qwen models accelerating iteration, weights opened to prioritize user scale; open-source monetization unproven. E-commerce: instant retail (Hema, Taobao Flash) losses narrowing faster than expected; AliExpress operationally profitable; cross-border revenue slight dip. Strategy shifts to balancing loss reduction with growth.

Baidu Enters AI-Native Growth Era

Baidu Q2 2026: core AI-driven revenue ¥12.5B, exceeding 50% of total revenue for second straight quarter. AI moves from concept to revenue base. AI applications: full office product matrix gaining user share, expanding from consumer to enterprise; leveraging Wenku, Netdisk, Search for model reuse. Monetization lags user growth. AI Cloud infrastructure revenue surging; GPU public cloud growth accelerating; internal compute standardized for external sale. Kunlun chips adapting mainstream models, building chip-model-cloud stack. Apollo Go (robotaxi) expanding globally, refining algorithms, fleet management, compliance. Valuation shifting to sum-of-parts; some overseas funds adding positions. Hong Kong secondary listing converting to dual-primary, aiming for Stock Connect inclusion to broaden investor base. Next phase: replicate commercialization model widely and stably.

WRC 2024: Full-Stack Robotics & $12.5k Humanoid

At World Robot Conference, Astribot (Stardust Intelligence) showcased a full-stack system: Lumo-2 foundation model for long-horizon reasoning, flexible object manipulation, millimeter precision — outperforming some overseas models on real tasks; DuoCore framework deploys large models on low-compute edge devices; AOS OS embeds AI into kernel scheduling, enabling voice-to-app generation for users, developers, researchers. T1 rope-driven humanoid mimics human tendon actuation, balancing dexterity and safety; debut price ¥89,900 (~$12.5k), drastically lowering entry barrier. Demos: 20+ complex chores, real-time air hockey, commercial reception/delivery. Industry trend: high-value embodied AI firms converging on vertical integration (model, OS, body) — akin to smartphone/auto evolution. Astribot, backed by ¥1B+ Series B, already deployed in pharmacy sorting, retail greeting, exhibition guiding, industrial material handling. Competition shifting from point-tech to system-level flywheels requiring co-evolution of model, OS, body in real-world data loops.

Expert Opinions

Fei-Fei Li: AI Fear, Spatial Intelligence, Governance

In a Bloomberg interview, Li critiqued extinction/AGI-dominance narratives as distractions from pragmatic policy. She acknowledged industry's communication gaps and concentration of technical power. Job displacement, she argues, strips dignity and agency; governance should foster broad societal participation in AI decisions, not halt progress. Li's World Labs bets on world models/spatial intelligence — language models lack physical perception for real-world embodiment. World models split into rendering, simulation, planning; Marble platform generates interactive 3D spaces for virtual production, game dev, robot sim. Still early; massive compute, talent, time needed. Humanoid robot investment huge but deployment gaps remain; spatial intelligence seen as key bridge. On regulation: reject utopian/doom binaries; guard against self-deification; risks (misinformation, misuse) mirror past tech revolutions; policy grounded in science, plus public R&D and STEM education. Optimistic on humanity's moral arc; calls for diverse voices in governance during this chaotic transition.

Tang Jie (Zhipu AI): Scaling Laws Beyond Parameter Count

Tang asserts scaling laws hold but parameter count alone is insufficient; must weigh training data volume, compute allocation, and deployment environment. History: early parameter races → Chinchilla (smaller model, massive data) → incorporation of inference cost → current models far exceed Chinchilla tokens-per-parameter. For MoE, total parameters = knowledge store; activated parameters = per-inference capacity; long-horizon tasks stress the latter. Zhipu's GLM-5.3 kept total/activated params fixed, concentrated compute on long-task environments and RL post-training, boosting complex benchmarks (internal tests, pending open weights). Demonstrates multiple scaling knobs; allocate compute to bottlenecks. Competition shifting: parameter count still references for simple QA, but for digital employees and long complex tasks, reliable end-to-end execution matters more. Trillion-parameter models not disappearing but no longer core marketing metric; optimal compute allocation is the persistent question.

Terence Tao & Wang Hong: Mathematics Must "Digest" AI Output

AI generates proofs faster than mathematicians can read. Both Fields medalists emphasize "digestion" — understanding, interpreting, integrating AI-generated proofs. Tao breaks research into five stages: generation, verification, exposition, community validation, knowledge integration. AI excels at first two; latter three remain human-dependent. Example: Sarnak conjecture — AI produced formally verified proof but unreadable; Tao spent days distilling core identity, compressing code, simplifying proof and covering a stronger conjecture. Digestion extends boundaries, not just error-checking. AI disrupts priority norms (journal/preprint timestamps vs. social-media drops skipping verification). Tao advocates elevating academic status of digestion work (exposition, refereeing); a machine-verified result without human-explainable insight should not be published. Launched Palomar registry: aggregates scattered AI proofs, records metadata, runs independent verification, serves as waystation between verification and publication; priority judged across generation, verification, exposition, publication. Sarnak proof first archived entry.

Technical Miscellany

SenseTime Open-Sources 8B SenseNova U1.5 Lite

SenseNova U1.5 Lite (8B params) supports native 4K output, matches closed-source GPT-Image-2 on complex layout, precise editing. Three weeks from preview to release; improved completeness, instruction accuracy, stability. Handles ultra-long complex prompts: multi-image stitching, local edits, bilingual text layout, multi-reference fusion; preserves composition, lighting, unaffected areas during local changes. Uses NEO-unify unified multimodal architecture (understanding, generation, editing in one model). Training: separate expert modules for text rendering, aesthetics, editing → multi-teacher online distillation into single 8B model, no routing overhead. Post-training reinforces instruction following, visual aesthetics, edit-region preservation; prompt-enhancement module tackles complex-task weaknesses (multi-reference decomposition, local text replacement). Evaluation shifting from visual wow-factor to end-to-end pipeline stability; unified model shortens call chain, lightens deployment, ensures cross-task consistency. Industry benchmark moving toward full-process performance: requirement understanding → generation → editing → delivery.

DeepMind's Recirculation: Deep-Layer Activation Feedback Beats Full Fine-Tuning

DeepMind proposes Recirculation: at inference, feed converged deep-layer activations back into shallower layers via a learned pathway, without modifying base weights. Standard Transformers pass information strictly forward; disambiguation often occurs late, leaving later tokens unable to reuse refined representations. Recirculation aligns norms of source (deep) and target (shallow) layers before fusion. On Gemma3, optimal layer pairs found per model size; works across open Transformers; benefit decays with token distance but observable at 256 tokens; verbs/adjectives gain most. Perplexity drops significantly; instruction following improves; downstream tasks mixed. Engineering: near-zero latency in autoregressive generation; noticeable prefill cost for long contexts. Adaptive Recirculation adds a tiny MLP controller predicting per-dimension recirculation coefficients from current token representation; base model frozen, only controller trained. On Gemma3 across nine language modeling sets, average perplexity reduction 23.0% vs. 21.6% for full fine-tuning; GSM8K math reasoning also gains. Controller sensitivity to training distribution limits universal stability. Shows performance gains possible via internal information-flow redesign, not just scaling or full fine-tuning; opens optimization avenue for frozen models.

Claude Disproves Century-Old Carathéodory Conjecture

Anthropic's Claude, with Harvard mathematician Levent Alpöge, constructed an explicit C∞ smooth convex surface with exactly one umbilic point of index 2, simultaneously falsifying Carathéodory conjecture (1924: every smooth closed convex surface in ℝ³ has ≥2 umbilics) and Loewner index conjecture (isolated umbilic index ≤1). Prior collaboration yielded counterexample to 3D Jacobian conjecture. AI role evolving from calculator to intuition partner, exploring spaces human intuition struggles to reach. Multiple classic conjectures falling via human-AI collaboration; new research paradigm likely to resolve more long-standing problems.