EAIDaily · 2026-08-03

Focus: AI Coding + Embodied Intelligence | Window: 2026-08-01 to 2026-08-03 Sources: OpenAI · Anthropic · Google DeepMind · GitHub · Microsoft · Supabase · Tsinghua · Unitree · Apptronik · Figure AI · AI HOT · Pondero · SiliconANGLE · TechTimes · dev.to

🎯 Headline Items (8)

AI Coding

1. OpenAI reveals Astra through ten machine-verified open-problem proofs — first non-sofic group in 27 years

What happened (Aug 1-2, 2026): OpenAI published a 249-page manuscript and a public GitHub repository containing Lean 4 machine-checkable certificates for ten mathematics and theoretical CS problems that had been open for at least a decade. The headline result is the first explicit construction of a non-sofic group — a question in group theory that has stood open since Mikhail Gromov introduced soficity in 1999. Other results include a disproof of Connes’s rigidity conjecture, a proof of Ehrhart’s volume conjecture, three Erdős problems closed (including problem 183 on multicolor Ramsey numbers), and the first improvement to the general upper bound on sphere-packing density in high dimensions since 1978.

Why it matters:

• Lean 4 certificates eliminate the trust layer. A Lean proof that type-checks is valid by construction; the GitHub repo reports a “sorry” count of zero. The “AI does math” debate moves from “is the proof correct?” to “is this the right problem?” — a fundamentally different conversation.
• Cost vs. research: ~$2,000 at Sol API rates for all ten solutions. The economics of attacking decade-old open problems now look like a cloud bill, not a sabbatical. Researcher Noam Brown noted on X that this was a deliberate ceiling, not a limit.
• Astra is OpenAI’s next major model family, designed for multi-agent long-horizon tasks (hours to days). Sam Altman demonstrated it in Washington D.C. — Astra will be the first system evaluated under the TRAINS pre-release government review framework.
• Math community reception is unusually strong: Thomas Bloom of erdosproblems.com called the results “big news” (ahead of OpenAI’s earlier Erdős unit distance counterexample); Terence Tao articulated a “big mathematics” vision of human-machine research division. The 2025 Leiden Declaration on AI/math crediting is already shaping the response.
• Open questions remain: Is it GPT-6 or a GPT-5 variant? No release date, no pricing. Peer review of the formal statements against community intent is the next gate.

Sources: SiliconANGLE · Pondero · AI Weekly · GitInformed

2. GitHub Copilot Workspace GA + Project Polaris at Microsoft Build 2026 — Microsoft swaps OpenAI for MAI under the hood

What happened (Aug 3, 2026): Two coordinated announcements at Build 2026:

• Project Polaris swaps the OpenAI model layer beneath GitHub Copilot for Microsoft’s internal MAI (Microsoft AI) model. Pro/Pro+ tiers gain multi-file context up to 100,000 lines and autonomous test generation by default. No price change, no new SKU — a margin decision wearing a product announcement. Every Copilot completion that previously routed through OpenAI’s API was a per-token toll on Microsoft’s own product.
• GitHub Copilot Workspace exits beta at general availability. Autopilot mode reasons across a full repository, proposes multi-file edits, runs tests, iterates autonomously, scoped by a GitHub issue. Fleet mode runs autopilot across multiple open issues simultaneously (dependency upgrades, style migrations, license sweeps). A new standalone GitHub Copilot desktop app surfaces all active agents in a “My Work” dashboard. Autonomous Agent Mode rolls out to Enterprise this month, with the human returning only at the final review-and-merge gate.

Why it matters:

• Microsoft is no longer the world’s largest OpenAI distribution channel — it is now a competitor. The biggest AI surface in software development no longer runs on rented brains. The competitive math against Claude Code and Cursor just got harder to model.
• The Workspace framing is structural: “the primary activity of a software engineer in 2027 will be reviewing and approving work done by agents, not writing that work.” That is no longer a prediction — it is a product surface.
• Sandboxing is the safety story for both modes. The unresolved gap, flagged by independent reviewers, is prompt-injection defense via repository context — the primary known attack vector against autonomous coding agents committing to production branches.
• xAI targets Grok 4.6 (~Aug 7, 1.5T params) and Grok 4.7 (2.1T params) weeks later. xAI is targeting a new foundation model every month through December, with Grok Build still iterating daily. The compression is across the whole sector.

Sources: dev.to / AI Daily Digest Aug 3 · Microsoft Build 2026

3. Anthropic launches Claude Tag — Slack-based multiplayer agent, 65% of Anthropic’s own code already agent-written

What happened (Aug 3, 2026): Anthropic launched Claude Tag, a Slack-based agent summonable by typing @claude in any channel. Now in beta for Claude Enterprise and Team customers; it replaces the older Claude in Slack app by Aug 3 with a 30-day admin opt-in. Claude Tag works with Opus 4.8 and turns Claude into a shared team agent with channel-scoped permissions, agent identity (service accounts, not borrowed user logins), separate memory between private workspaces, audit logs, and admin controls. Launch credits: $25,000 (Enterprise) / $2,500 (Team ≥10 seats), expiring Sep 1.

Why it matters:

• 65% of Anthropic’s own product team’s code is created by its internal version of Claude Tag. This is the strongest evidence yet that a frontier lab’s internal operating model is now AI-native end-to-end, and Anthropic is selling the same harness externally.
• The “agent identity” model is the real architectural move. Channel-scoped permissions + service accounts + org-billed channel work vs. user-billed DMs is a billing, security, and governance primitive — not a chat feature.
• Multiplayer AI work is the new category. This positions Claude Tag as an evolution of Claude Code and Claude Cowork into shared workflows. Combined with GitHub Workspace GA, the workspace is becoming the agent’s home, not the IDE.
• Per-channel billing is a clever wedge. It lets Anthropic grow with team activity rather than per-seat, and it monetizes the moments when agents actually produce value, not when humans are typing.

Sources: wpnews.pro / Testingcatalog · Anthropic

4. Supabase open-sources Evals framework — Claude Code skips 60%+ of real DB documentation checks

What happened (Aug 1, 2026): Supabase released Supabase Evals under Apache-2.0 — an open-source coding-agent benchmark that spins up a real Supabase environment and runs agents against actual tasks: building a schema, debugging a failed Edge Function, fixing a broken Row Level Security policy. Claude Code, OpenAI Codex, and OpenCode are the three agents in the initial release. Each scenario uses real MCP server calls and the Supabase CLI rather than synthetic fixtures. Results feed a public leaderboard at supabase.com/evals.

Headline finding: Claude Code skips documentation checks in over 60% of real database tasks, relying on pre-training knowledge that may be months out of date.

Why it matters:

• Production-grade tasks are the new evaluation frontier. Schema drift and live config state matter; toy-problem benchmarks don’t. Supabase Evals complements Sentry, Vercel, and the new MLPerf coding evals as a real-world signal.
• The “harness > model” thesis now has its own measurement infrastructure. A leaderboard where the agent choice and the harness are explicitly co-tested is a procurement-grade artifact, not a benchmark for bragging rights.
• The 60% documentation skip is a real product gap. Frontier agents optimize for “the code works” — but production engineering requires “the code works and matches the current docs/spec/version.” Most evals can’t see that. Supabase Evals can.

Sources: Pondero daily brief Aug 2 · Supabase blog · Tech Insider comparison

5. Tsinghua VeriLoop Coder-E1 open-sources “evidence-spiral” — 85.20 SWE-bench Verified at 27B

What happened (Aug 2, 2026): The Tsinghua Shenzhen International Graduate School Intelligent Robotics Lab (PI: Prof. Liu Houde, AI lead: Dr. Wang Libo) released VeriLoop Coder-E1, a 27B-class code model built on Qwen3.6-27B for repo-level code repair and agentic software engineering. Results on Hugging Face’s benchmarks (as of Jul 27):

• SWE-bench Verified: 85.20 (#1 in ≤32B open-weight)
• SWE-bench Pro: 62.38 (#1 in ≤32B open-weight)
• Terminal-Bench 2.0: 76.40 (#1 in ≤32B open-weight)
• DeepSWE: 33.63 (#2 in ≤32B open-weight)

The model is a narrow-domain PEFT fine-tune with a detachable Surface Host Adapter — original Qwen3.6-27B weights are frozen; a small adapter reinforces tool-contract compliance, evidence-conclusion binding, uncertainty recognition, validation-failure explanation, and patch/rollback boundary control.

Why it matters:

• An evidence-spiral paradigm (verifiable recursive self-improvement) is a third path between “bigger model” and “bigger dataset.” The narrow PEFT strategy preserves the base model’s general capability while pushing the agentic coding needle into frontier territory at one-tenth the cost.
• Top of every major SWE benchmark in its size class, including beating some larger open-weight models. It places the academic + small-team open-weight track on the same league board as Kimi K3 (2.8T) and Claude Sonnet 5 (proprietary).
• The verifiability framing is the bigger idea. A self-improving coding agent that requires its improvements to be evidence-grounded — tool contract compliance, validation failure explanation, rollback boundaries — addresses the long-horizon “agent drift” problem that SlopCodeBench exposed.

Sources: qq.com / 清华团队 VeriLoop · Tsinghua iVG Lab

Embodied Intelligence

6. Unitree H1 Pro three-continent launch + STAR Market IPO sprint — $250M revenue, 60.27% gross margin, Pentagon-listed

What happened (Aug 1-12, 2026): Two threads converging on the same week:

• Commercial rollout: Unitree launches the H1 Pro humanoid commercially in Europe (Jul 22), Asia (Aug 5), and North America (Aug 12) — the first coordinated three-market humanoid launch the industry has attempted. Boston Dynamics, Tesla, Figure, Agility, and 1X have never tried it. 500+ H1 Pro units will deploy to Asian logistics and manufacturing at launch partners.
• Capital markets: Unitree’s STAR Market IPO enters the final stretch. Preliminary bookbuilding Aug 5; online/offline subscription Aug 10. H1 2026 revenue guidance: ¥1.052–1.128B (~$155M), +35.62%–45.41% YoY. 2025 humanoid shipments: 5,500+ units, #1 globally (Unitree’s prospectus claim), $250M revenue, 60.27% gross margin. Hong Kong-listed peers rallied on the news: Xiangong +22%, Lifu Harmonic +16%, Geek+ +5%.

The regulatory backdrop, all hitting the same week:

• Pentagon Section 1260H list (Jun 8) designates Unitree hardware as Chinese military technology.
• GUARD Act (H.R. 9129, Moolenaar/Obernolte/McClellan) routes Chinese humanoids toward the FCC’s Covered List.
• EU AI Act Article 50 (Aug 2) brings the chatbot disclosure and deepfake labeling regime live.
• China’s National Intelligence Law Article 7 obligates Unitree robots to cooperate with state intelligence regardless of where they operate.

Why it matters:

• Three-continent launch + IPO + blacklist in one quarter is the new bar for the industry. Hardware unit-economics (G1 BoM ~$8,976 vs. $13,500 list) plus the EV supply-chain actuator base makes the price floor defensible.
• R&D reality check: During the first 9 months of 2025, research/education was 74% of humanoid revenue; industrial customers were 9% (and half of that was tours, not labor). The “real labor at scale” thesis is still unproven even for Unitree.
• Hong Kong market is treating Chinese humanoid as a capital-markets scarcity play amid cooling global AI speculation. UBTECH, AgiBot, and Deep Robotics listings compound the pressure.
• Compliance burden sits with the buyer, not the maker. EU AI Act high-risk classifications put conformity assessment, human oversight, and documentation on whoever deploys. That is a procurement team problem, not a Unitree problem.

Sources: TechTimes · Frontiernews · Caixin via Futu · ai2.work / HumanoidApplications.com

7. AXIS data engine proves crowdsourced sim can train production robot policies — 153M+ trajectories, π0.5 +4.9pp

What happened (Aug 2, 2026): Axis Robotics, together with Georgia Tech, UC Berkeley, and other universities, released the AXIS community-driven data engine and evaluation benchmark for robot manipulation. The dataset, built from web crowdsourced teleoperation trajectories in simulation, reached 1.53M+ trajectories by July 24. Using π0.5 as the base model, full AXIS data lifts LIBERO-Plus overall success rate by +4.9 percentage points over baseline.

Why it matters:

• This is the first rigorous proof that crowdsourced sim data can train a real VLA. Until now the assumption was that high-quality manipulation data had to come from in-house teleop (DoorDash, Figure+Brookfield, Tesla towers, China government training centers). AXIS shows a path to data flywheel scale that looks more like ImageNet than like Figure Helix.
• The data collection cost curve bends. Teleop data fell from ~$340/hr (2024) to ~$118/hr (2026). Crowdsourced web sim may bend it again.
• The “embodied data drought” narrative has a credible counter-strategy. Combined with SmolVLA (450M params) and NanoVLA (52× faster inference), the small-model + small-budget + open-data stack is now a coherent alternative to the Figure/Apptronik/Tesla full-stack.
• Implication for unit economics: A +4.9pp delta on LIBERO-Plus, at 1.53M trajectories, suggests the marginal value of additional trajectories is still positive — the scaling law for crowdsourced data is not yet saturated.

Sources: 机器之心 / 36kr ecosystem · qq.com GitHub日报 · Axis Robotics

8. Chengdu consolidates 5 of China’s humanoid giants — 6,400+ robots deployed by Nov, 10K by year-end

What happened (Aug 2, 2026): Unitree, UBTECH, AgiBot (Zhiyuan), Galaxy General (Yinhe Tongyong), and Stardust Intelligence (Xingdong Jiyuan) — five of China’s top humanoid players — have now all committed to Chengdu. Unitree’s GD01 mass-produced manned变形机甲 (transformable mecha) had its global debut in Chengdu on Jul 24, paired with a city-government strategic cooperation agreement.

Deployment targets from the local government’s real-scene training plan:

• 60+ real-scene training sites active
• 6,400+ whole robots deployed by November 2026
• Target: 10,000+ by year-end

Why it matters:

• A single Chinese city is becoming the equivalent of Silicon Valley + Detroit for humanoids. Five top-10 players co-locating creates a supply-chain + customer + talent + policy flywheel that is hard for any Western hub to match. Suzhou already runs the National and Local Co-Built Humanoid Robotics Innovation Center; Shanghai has Geek+ warehouses; Shenzhen has the consumer side. Chengdu is now the integrated mass-deployment hub.
• The “ten-thousand-robot deployment” threshold is the embodied AI equivalent of a million-server cluster. At that scale, integration logistics, after-sales service, and data flywheel all start compounding in the same geography.
• Beijing’s 2026 target is 100,000+ humanoid units nationally. The 5-giant Chengdu cluster would represent ~10% of national output from a single city.
• Cross-check on the commercial thesis: Geek+ warehouses + StarDust’s sortation in Shunfeng/China Post logistics centers + Unitree’s industrial pilots = real revenue from real labor, not just demos. The Chinese humanoid industry is the first to break the “lab-experiment” framing at scale.

Sources: qq.com / 具身智能日报 Aug 2 · 科创四川 · China Industrial News

📊 Quick Takes (6)

🧭 Trend Lines (5)

1. AI math moves from “trust the lab” to “verify the proof.” OpenAI Astra’s Lean 4 certificates make ten open problems machine-checkable for $2,000. The “AI does math” PR cycle is replaced by a public-leaderboard-of-proofs cycle. What changes for engineering: the same formal-verification loop becomes feasible for code, configs, and infra — agent outputs may soon ship with machine-checkable guarantees, not just test passes.

2. Microsoft ends its OpenAI distribution deal — in code, not contract. Project Polaris puts MAI under Copilot, the largest AI surface in software dev. OpenAI’s distribution leverage just shrunk; Anthropic’s Slack-based Claude Tag (65% of Anthropic’s own code) shows the next distribution surface is team chat, not the IDE. What changes for procurement: enterprise AI stacks will increasingly look like multi-model + agent-identity + workspace-first, not single-vendor.

3. The “harness layer” is now a product category. YC’s QM, Anthropic’s Claude Tag, GitHub Workspace Autopilot/Fleet, Supabase Evals, AMD+CMU’s FlashRT (70× lower latency on B200 for multi-GPU coding-agent deployment), and CodexLoom (long-term Agent teams from Codex sessions) all shipped in the same 72 hours. What changes for builders: the model choice is now a downstream decision of the harness choice, not the other way around.

4. Open-weight coding models are no longer “fast followers” — they lead specific size classes. VeriLoop Coder-E1 (27B) leads SWE-bench Verified/Pro/Terminal-Bench 2.0 in its class; DeepSeek-V4-Flash matches Opus 4.8 agent capability at 40% of the cost; Kimi K3 is the first 3T-class open-weight model with self-hostable weights. What changes for buyers: any coding workload that doesn’t require GPT-5.6 Sol / Claude Opus 5 will route to an open-weight default by Q4 2026.

5. Embodied AI is consolidating into a few global deployment hubs. Chengdu (5 giants, 6,400+ robots by Nov), Austin (Apptronik’s 90K sqft “Robot Park”), Spartanburg (Figure 03 at BMW), Suzhou (National Humanoid Innovation Center), Memphis (xAI’s 200K-GPU Colossus), San Jose (Figure + NVIDIA + OpenAI). What changes for supply chains: the embodied-AI supply chain is concentrating faster than the model layer; whoever controls the regional deployment cluster controls the data flywheel.

📈 Benchmark Snapshot (Aug 1-3, 2026)

Coding Agents (model + harness combinations)

Embodied Intelligence

Frontier Model Context (Aug 1-3)

🪞 What to watch this week (Aug 4-10, 2026)

• Aug 5: Unitree STAR Market IPO bookbuilding opens; Unitree H1 Pro Asia commercial launch; Honda × Sony humanoid JV announcement; Sony Vision-S commercial release
• Aug 7: xAI Grok 4.6 expected release (~1.5T params, monthly cadence target)
• Aug 10: Unitree STAR Market IPO subscription begins; Humanoids 2026 IEEE-RAS conference in Genoa
• Aug 12: Unitree H1 Pro North America commercial launch
• Aug 14: Honda Asimo successor E4 begins hospital/elderscare field testing in Japan
• Open question: Does OpenAI confirm Astra as GPT-6 or as a GPT-5 variant? Does the TRAINS pre-release review set a template?
• Open question: Does the EU AI Office formally request OpenAI’s training-data summary under the copyright chapter? Does Brussels escalate the rogue-agent informal talks to formal proceedings?
• Open question: Does Microsoft extend MAI under Copilot to GitHub Workspace, or keep MAI as a Copilot-only model?

EAIDaily is an automated English daily digest focused on AI Coding and Embodied Intelligence. Sources are linked inline; benchmarks reflect publicly disclosed results and may shift with subsequent updates. Trends are working hypotheses, not predictions.