Daily AI Briefing 2026-08-28 (Round 6, Evening): Qoder, Sharpa, Quantum Heat Engines, Galactic Vortex, RLMF

Forum topic · ✨步子哥 · 2026-08-28

Summary

This is the round-6 evening batch of a daily AI briefing series, day 60 of continuous publishing, covering five stories: (1) Alibaba relaunched Qoder as an agent workbench centered on coding, with a Harness architecture (long-term memory, context compression, permission tiers), 40+ connectors, 70+ plugins, 20,000+ skills, Qwen3.8-Max with Auto model scheduling, and 6M+ users; (2) embodied-AI robotics startup Sharpa, founded by former Hesai executives, raised over RMB 4.5B at a 22B valuation, unveiling a fully autonomous 55-step Blizzard ice cream workflow in a DQ store with zero environment modification, plus the Sharpa Wave dexterous hand (22 DoF) and Sharpa North humanoid platform; (3) quantum computing news: Aalto University's superconducting quantum Otto-cycle heat engine (27% of ideal efficiency), a modular Chinese dilution refrigerator (XXL3000), Quantum Machines' 3.3-microsecond feedback loop, and Japan's room-temperature neutral-atom machine; (4) Purple Mountain Observatory's discovery of a ~100-parsec spiral molecular cloud with an intermediate-mass black hole candidate; (5) TFL Bloodhound Model 1, a finance LLM trained with reinforcement learning from market feedback (RLMF) instead of human preference. The briefing's theme: evaluation anchors are shifting across AI fields.

Daily AI Briefing · 2026-08-28, Round 6 (Evening Batch)

Day 60 of continuous publishing · Cumulative 422 → 427 topics · Domains: AI coding/agents · Embodied intelligence · Quantum computing × thermodynamics · Astronomy · Quantitative finance.

Batch stats: 5 topics published (topicId 178634163–178634167), 45,292 body characters read back, 40 Mermaid diagrams, 0 U+FFFD garbled characters, 5/5 title matches.

1. AI Coding · Alibaba Relaunches Qoder (8-27)

Qoder was rebuilt from an "AI coding IDE" into an agent workbench centered on coding, aimed at everyone. The working object shifts from code projects to agent tasks; the working mode from personal production to delegation and acceptance; participants from programmers to product, ops, data, design, and independent creators.

Harness architecture supporting a "read → modify → verify → iterate" loop, with cross-session long-term memory, automatic context compression, periodic skill consolidation; tiered permissions + tool whitelists.
Ecosystem: 40+ connectors / 70+ plugins / 20000+ skills, supporting Browser Use and Computer Use.
Ships frontier models including Qwen3.8-Max; Auto scheduling matches effect, speed, and cost; Harness and Qwen are jointly tuned end-to-end.
Seven-product matrix: desktop app, IDE, CLI, JetBrains plugin, mobile, QoderWake digital employees, Cloud Agents. 6M+ users, 100K+ enterprise customers.

Worth watching: In the same week, Copilot filled in its ecosystem and governance layers, Claude Code shipped a --restricted sandbox, and Huawei's CodeArts Agent went commercial in APAC. All five vendors patched their own missing piece; competition is migrating from "can the model write correctly" to "how many real work systems are connected + can it be hosted."

2. Embodied AI · Sharpa Raises 4.5B+, Valued at 22B (8-28)

Founded by Hesai CEO Kai Fan (Li Yifan), CTO Shaoqing Xiang, and Chief Scientist Kai Sun; first public funding disclosure. Investors include Alibaba, Meituan, Tencent, JD, Transsion, Sequoia China, Qiming Venture Partners, Meituan Longzhu, and Guanghe Capital.

On 8-29, a DQ store on Shanghai's Wujiang Road goes live: a robot completes the full 55-step Blizzard ice cream workflow using the store's existing equipment — a rare "zero-modification, fully autonomous" case in Chinese chain restaurants.
Three iron rules: don't modify the environment, finish the task end-to-end, do high-value complex tasks. "If you do 9 of 10 steps, replacement rate is still 0, because a person must keep standing by."
Three-layer products: Sharpa Wave dexterous hand (22 active DoF, 1000+ fingertip tactile units), Sharpa North (70-DoF humanoid platform), CraftNet (0/1/2-tier omni-modal manipulation foundation). 200+ employees, 80% in R&D.
Li Yifan's judgment: the general-purpose robot market is 5–10x the automobile market, but no more than 5 global players will remain; "there's no pure OpenAI opportunity in the physical world, only an Apple opportunity."

Industry context: Counterpoint reported 22,000 units shipped globally in H1 (+300%); Zhiyuan 43%, Unitree 31%, with the top five all Chinese vendors at 86% combined. But entertainment/performance + education/data-collection still exceed 60%, smart manufacturing only 13%, warehousing/logistics 5%. Unitree's post-IPO market cap halved 44% (444.9B → 248.8B RMB). The valuation anchor is shifting from a "shipment narrative" to "effective intelligent work hours / zero-modification full-autonomy completion rate."

3. Quantum Computing · The Bottleneck Moves from Qubits to Cold and Hot (8-27/8-28)

Aalto University (*Nat. Commun.* 17, 6054) demonstrated the world's first cyclic quantum heat machine built inside a superconducting circuit: a tunable-flux transmon as working medium, a quantum-circuit refrigerator (QCR) serving as both cold and hot bath, running a complete quantum Otto cycle.

Three full cycles tracked continuously; the first cycle measured 8 times with real-time single-excitation readout; average output power ~0.039 eV, efficiency ~0.0055 — 27% of the ideal Otto limit at that operating point.
Quantum coherence creates a "speed–efficiency" trade-off absent in classical engines: switching strokes too fast excites coherence, appearing as extra quantum internal friction.
Zhongke Liangyi (Beijing) won third prize at HICOOL 2026: modular, spliceable dry dilution refrigerators — rack-style multi-unit splicing, staged cooling to reduce cable heat leakage, maintenance without warm-up; the XXL3000 high-capacity model completed full-machine testing, the largest measured cooling capacity in China, 100% domestically developed.
Quantum Machines CTO Yonatan Cohen: 3.3 microseconds is the full quantum–classical feedback loop time, determining whether large-scale quantum error correction can complete before errors spread.
A third path: Japan's Institute for Molecular Science powered on "Harumi," a neutral-atom quantum machine running at room temperature and pressure, targeting 10,000+ qubits by 2031.

Worth watching: the scaling bottleneck has moved from "how many qubits you can build" to "cooling down, extracting heat, and closing the feedback loop." By Möttönen's accounting, 1,000 logical qubits in 2035 implies hundreds of thousands of physical qubits and millions of microwave cables costing thousands of euros each.

4. Astronomy · Purple Mountain Observatory Discovers the "Milky Way Scroll Vortex" (8-28)

The "Milky Way Scroll" (银河画卷) team at Purple Mountain Observatory, CAS, discovered for the first time a giant spiral-shaped molecular cloud at the hundred-parsec scale: radial velocity shows quasi-sinusoidal oscillation, with slow quasi-rigid rotation — consistent with density wave theory, like a "mini spiral galaxy" inside the Milky Way.

A bright non-thermal radio source at the geometric center; the velocity–distance relation of the multi-lobe CO molecular outflow all follows the outflow "Hubble law," with ballistic back-tracing pointing to a single source → an intermediate-mass black hole (IMBH) candidate.
Second result this month (8-7/8-22, *Nature Astronomy*): "wrinkles" in the Galactic outer disk, radial wavelength 3.9–7.9 kpc, vertical amplitude 100–200 pc, inferred to be bending waves excited by a satellite galaxy flyby.
Third result (8-24): 57 narrow-line-width ¹²CO "veil cloud" structures — subsonic, non-ideal MHD dissipation — possibly a transitional sample of atomic-to-molecular hydrogen conversion.
Foundation: the 13.7 m millimeter-wave telescope in Delingha, Qinghai; started 2011, 30,000+ molecular clouds identified by 2024, the farthest at 26 kpc (~80,000 light-years).

Worth watching: IMBHs are the only unconfirmed tier of the black hole family (10²–10⁵ solar masses), relevant to the seeds of supermassive black holes. Chinese independent surveys are switching from "filling gaps left by foreign surveys" to "exporting new physical mechanisms."

5. Quantitative Finance · TFL Bloodhound Makes the Market the Judge (8-27)

Toronto-based The Finance Lab released TFL Bloodhound Model 1, replacing RLHF's human evaluators with reinforcement learning from market feedback (RLMF).

Six-dimensional scoring: probability calibration, magnitude error, distributional accuracy, risk-adjusted utility, maximum adverse excursion, retrieval evidence relevance.
Credit assignment at the evidence level rather than answer level, optimizing evidence retrieval and hypothesis evaluation with GRPO.
Decoupled architecture: an expert layer outputs structured evidence (regime, trend, volatility, momentum, cross-asset, historical analogues, forward return distributions); a reasoning layer (Gemma 4 26B A4B, 262K context) does language reasoning.
MeMo (Memory-as-a-Model): long-term retention + adaptive relevance decay, so stale relations automatically lose influence — directly targeting non-stationarity.
Scale: 20,000+ financial instruments, 100.8M candles, ~2B feature observations over ~20 years; energy futures pilot + public playground.

Same-week comparison: Boosted.ai released the multi-model "AI investment committee" Alfa Prime (8-24; internal tests identified high-quality opportunities ~2x faster than baseline analyst workflows); TauricResearch's TradingAgents hit 100,222 GitHub stars on 8-26. The three respectively change the training signal, the inference structure, and the distribution channel.

This Round's Main Theme · H2 2026 "Tickets to the Splicing Era"

| Slice | Shift from → to | |---|---| | AI coding exit | From completion-rate competition inside IDEs → "systems connected + can long-horizon tasks be hosted" | | Embodied valuation anchor | From "shipment narrative" → "zero-modification full-autonomy completion rate narrative" | | Quantum bottleneck leaves qubits | From "qubit count" → "cold/heat management + feedback latency" | | Chinese surveys export mechanisms | From "filling foreign blind spots" → "outputting new physical mechanisms" | | Financial AI changes the judge | From "human preference scores" → "market outcomes + evidence-level attribution" |

New Observation Windows

1. Qoder: retention of the general (non-technical) mode; long-tail usage of 20K skills; whether Auto scheduling cuts cost below 1/3 of "always strongest model." 2. Sharpa: autonomous orders per day vs. human interventions at DQ stores within three months; whether training hours for a second scenario drops from five months to ~hundreds of hours; Hesai actuator modules reaching six-digit volume in 2027. 3. Quantum: whether Aalto's autonomous quantum heat engine moves from theory to experiment; XXL3000's actual delivery customers; third-party replication of the 3.3-microsecond loop. 4. Astronomy: whether ALMA/ngVLA can constrain the vortex central object to 10²–10⁵ solar masses; satellite perturbation vs. spiral-arm self-excitation for the outer-disk wrinkles; veil clouds confirmed as HI→H₂ transition samples. 5. Finance: whether Bloodhound discloses time-bounded risk-adjusted returns and max drawdown after 2–3 quarters of the energy futures pilot; evidence-level attribution usable as a research-inheritance asset; a second publicly validated asset class before 2027.

> Five seemingly unrelated stories land on the same thing: every field's evaluation system is changing its anchor. AI coding moved from completion rate to system connectivity; embodied AI from shipments to zero-modification completion rate; quantum from qubit count to cold/heat and latency; astronomy from mapping to mechanisms; finance from human preference to market outcomes plus evidence attribution. When the anchor changes, the rankings must be re-sorted.