LectureRechercheRadarCadre d'investissement
Connexion / Inscription
ToutTechnologieProduitsInvestissementBourseAutres

Lecture

2026-11-0499 publications

Actualiser

Wang Hao, Founder of Yongquan Innovation: Five-Axis Lacks PMF Today; Three-Axis is the Correct Starting Point for Desktop CNC

Sujet · 涌泉创新桌面CNC路线Matériel
2026-11-04 10:54 GMT+8

The 2026 desktop CNC sector is being ignited by the concept of five-axis machining, with multiple players launching crowdfunding campaigns centered on five-axis as their core selling point, seemingly suggesting that whoever launches five-axis first gains the first-mover advantage. Wang Hao, founder of Yongquan Innovation, holds the exact opposite view: five-axis currently lacks Product-Market Fit (PMF); only a very small number of items absolutely require five-axis processing, while the majority of needs can be satisfied with three-axis.

Wang Hao's reasoning is built upon his long-term research into the Maker industry. After leaving DJI in August 2023, he spent over a year studying the entire desktop manufacturing landscape, covering metal 3D printing, multi-material printing, UV printers, and CNC. He almost exhaustively reviewed user discussions on Reddit and Discord, manually performed business intelligence analysis, gathered nearly all available data regarding desktop CNC product demand online, and then integrated this with his own practical needs as an engineer to define the product direction.

After defining the product, he conducted quantitative validation through questionnaires and user interviews, finalizing parameters item by item. Just as the product definition was taking shape, Chaojiang Technology launched its first desktop CNC, whose product definition was quite similar to what Wang Hao had envisioned at the time. With the first-mover already occupying the space, Wang Hao re-evaluated: the direction remains unchanged, but he needed to figure out how to do it better.

He identified three main differentiators: a target processing power of 1.5 kW, which is 100% higher than Chaojiang Technology's; a processing area of 400×300 mm, offering three times the volume; and a design centered on three-axis capability but retaining the ability to upgrade to five-axis. A North American user wanted to modify motorcycle parts requiring a total length of approximately 16 inches and worried that an 800W spindle would lead to low efficiency—this user's needs perfectly aligned with Wang Hao's three key points.

Regarding five-axis, Wang Hao's logic is that while five-axis will certainly meet future user needs, it currently serves only a very limited number of scenarios. Many manufacturers making five-axis machines have habitual thinking, assuming that since Bambu Lab makes sculptures, CNC should also make sculptures. However, the primary use case for five-axis today is not sculpture—four-axis handles sculpture well, and 3D printing does so too; with so many competing processes, five-axis offers little advantage. Many competitors initially defined their products around AI-generated sculptures followed by machining, a concept carrying PMF risk; users are not buying into it, as they primarily want to make components.

Yongquan Innovation's crowdfunding results support this judgment. Having been established for just over a year, the company was not among the earliest crowdfunding players this year, yet it entered the public eye with over $16 million in new product crowdfunding. Wang Hao believes that Chaojiang Technology and Zao Wu Shi Dai, despite having correct product definitions, achieved decent crowdfunding results even if their implementation wasn't extreme; Yongquan performed better in product definition, implementation, and marketing, hence achieving superior results.

Sources :leiphone.com

Recherche

Generative Media Inference Startup fal Raises $1.3B at $16B Valuation

Sujet · fal融资动态Matériel
2026-11-04 06:32 GMT+8

Generative media inference startup fal announced on November 3, 2026, that it has raised $1.3 billion in new funding at a valuation of $16 billion. According to Bloomberg, the round was led by Bessemer Venture Partners, with participation from Nvidia, Sound Ventures, HOF Capital, and T. Rowe Price Group. Bessemer partner Talia Goldberg will join fal's board.

Fal provides a cloud platform for running image, video, and audio generation models, accessible via a single API. Its customers include entertainment studios, advertising agencies, and game developers. Co-founder and CEO Burkay Gur stated that the large round reflects surging demand and the need for more compute power. He noted that the company had to rate-limit customers during the summer because it could not serve all requests.

The announcement also covered an April Series E round led by Sequoia Capital, which raised $350 million at an $8 billion valuation. Combined with a $140 million raise last December, fal has now secured over $1.9 billion in total funding. A person familiar with the matter told Bloomberg that fal's annualized revenue passed $800 million earlier this year.

This valuation places fal just behind rival inference provider Fireworks AI Inc., which closed a $1.5 billion round at a $17.5 billion valuation in July. Several people familiar with the finances of inference companies told TechCrunch that gross margins across the sector remain thin due to the high cost of buying or leasing compute. Gur said pretraining a model of fal's own is not on the near-term agenda but did not rule it out for later.

Sources :SiliconANGLE

Recherche

Apple Opens Siri AI Integration to All Developers, Adds 'Skills' for Multi-Step Tasks

Sujet · 苹果Siri技能整合Matériel
2026-11-04 11:10 GMT+8

Apple has announced that it is opening its next-generation Siri AI integration to all third-party developers, accompanied by a new mechanism called "Skills." Released as part of the iOS 27.3, iPadOS 27.3, macOS 27.3, and visionOS 27.3 betas, this update marks a shift for Siri from executing single commands to acting as an AI agent capable of handling complex tasks across multiple applications.

Previously, Siri's interaction with third-party apps relied heavily on Apple-defined schemas or limited Shortcuts. Now, developers can provide broader custom operations through the App Intents framework, while the new "Skills" mechanism allows Siri to chain multiple actions into a complete workflow. For example, a user can ask Siri to find a saved item in a shopping app, identify a specific contact, and send the product details via a messaging app, all without manually switching between interfaces.

This move aligns with Google's recent push for Android AppFunctions and Gemini's multi-step task capabilities, as both mobile ecosystems compete to establish their AI assistants as unified entry points for service invocation. For developers, this implies that beyond designing intuitive human-facing interfaces, they must also consider how to make their services understandable and efficiently callable by AI agents.

It is important to note that these features are currently in developer preview and are expected to reach general users in the coming months, with App Store submission channels opening in spring 2027. Additionally, Siri AI does not yet officially support Traditional Chinese, and its real-world utility will depend on the depth of third-party integration and operational reliability.

Sources :TechOrange 科技報橘

Recherche

Salesforce AI Research: Full-Path Evaluation Distinguishes 'Failed to Store' from 'Failed to Use'; Agentforce Memory Strategy Adjustment Doubles Recall

Matériel
2026-11-04 08:18 GMT+8

When an AI agent gives an incorrect answer, developers face a difficult diagnostic problem: was the error because the agent never saved that information in the first place, or because it saved it but failed to retrieve it when needed? Erik Nijkamp, Director of Research at Salesforce AI Research, noted in a blog post that one cannot distinguish between these two failures based solely on the agent's final response, because a reply generated based on outdated facts reads exactly the same as one generated based on current facts.

To address this diagnostic blind spot, Salesforce AI Research built an evaluation framework: it generates long conversations spanning multiple sessions, plants specific facts within the dialogue, feeds the complete conversation into the running Agentforce system, and then asks questions that depend on those facts to verify whether the agent answers correctly. The key to this method lies in testing the entire path—from 'the customer stating the information' to 'the agent acting on the information'—rather than checking only the storage layer or only the retrieval layer. The conversational data used for the evaluation comes from ConvoMem, a public benchmark released by the team in 2025.

Based on this full-path evaluation, the team found that improvement did not require touching the architecture. The underlying architecture of Agentforce Memory has remained unchanged since its launch; what actually changed was the memory policy—the rules determining which customer information the agent retains and how it handles changes to that information. The old policy treated a customer's update as a new fact, storing it alongside the old fact; the new policy treats the update as a correction to an existing fact, overwriting the old value.

Evaluation results run on their own system show that the new policy achieves 1.8 times the recall of the old policy for explicitly stated facts, and 2.2 times the recall for facts that change over time. Salesforce summarized this as end-to-end conversational memory quality being 'essentially doubled.'

It is important to clarify that all the above figures are self-test results conducted by the vendor on its own product using its own published benchmark, with no third-party independent replication yet available. The evaluation did not disclose absolute recall rates, conversation sample sizes, or statistical confidence intervals. Therefore, while the multipliers of 1.8x and 2.2x indicate the directional effect of the strategy adjustment, they are insufficient to judge the absolute performance level in actual customer deployment scenarios.

Nevertheless, this methodology itself holds transferable value: when an agent's memory performance is poor, establishing a full-path evaluation first to distinguish between storage failure and retrieval failure, and then targeting strategy modifications to the corresponding layer, is often more efficient than rebuilding the architecture. Silvio Savarese, Chief Scientist at Salesforce, summarized this approach as the difference between a product organization and a laboratory—the problem is already running, customers are already using it, and the discipline is to measure honestly before making any changes.

Sources :Salesforce

Recherche

Falk Steiner, Digital Policy Reporter at heise online: The Real Risk of European Sovereign Cloud Is Not Data Access, but Software Update Supply Cutoff

Sujet · 欧洲主权云依赖风险Matériel
2026-11-04 01:38 GMT+8

At the Berlin Digital Summit, German Chancellor Friedrich Merz was asked whether it sends a bad signal to domestic data center operators if German intelligence agencies rely on Google Cloud for their tasks. His answer was negative, noting that Google also invests in Europe. He expressed high confidence in Germany's intelligence services, stating that even when an American provider builds large cloud infrastructure, the data remains secure.

Falk Steiner, digital policy reporter at heise online, argues that this assessment fundamentally misidentifies the problem. In a commentary, he points out that the scenario requiring genuine concern is not Google handing over German data to US intelligence agencies under presidential directive—a situation that can indeed be prevented through technical and organizational measures within an air-gapped cloud architecture. The actual danger lies in a different scenario: a soft supply cutoff switch.

Steiner lists all current cloud products marketed as European sovereignty solutions—Amazon’s European Sovereign Cloud, the Microsoft Azure-based Delos Cloud, Google’s partnership S3NS in France, and the planned Thales Cloud in Germany. Technologically, they all depend on the maintainer of the cloud solution's software stack. The software stack is a core criterion for choosing or rejecting a cloud vendor, second only to cost.

The issue is business continuity: if a displeased US President orders a halt to providing software updates to third parties, these so-called sovereign clouds will begin to show cracks within a few months. At that point, migrating data and cloud applications to a functional target infrastructure would be necessary—but that target infrastructure must first exist.

Here lies a deadlock: without anchor customers (such as German intelligence agencies with the highest security requirements), there is no economic incentive to develop and operate an independent infrastructure capable of replacing US solutions. Without this infrastructure, intelligence agencies would have nowhere to migrate when a supply cutoff occurs.

At the EU level, the 'Cloud and AI Development Act' (CADA) is currently being drafted to set relevant standards for sovereignty; Germany also plans to ensure independence across the entire public sector and among commissioned private service providers. Steiner argues that in this context, Germany’s decision to bet its intelligence agencies’ most classified knowledge on a Google Cloud solution constitutes a betrayal of the entire digital sovereignty effort.

Steiner concludes that when a real choice must be made, sovereignty takes a back seat. Those who recently pledged billions of euros to Chancellor Merz at an investment summit, hoping Germany would drive Europe toward greater independence, will remember this signal.

Sources :heise online

Recherche

Alibaba Cloud Private Cloud Internationalization: <5% Revenue Share Consumes Nearly Half of Resources, Executive Says It's a Long-Term Play

Matériel
2026-11-03 23:20 GMT+8

According to a report by Leiphone, Liu Guohua, Vice President of Alibaba Cloud Intelligence Group and General Manager of Private Cloud, recently disclosed the latest financial and resource allocation status of Alibaba Cloud's private cloud international business. The data shows that although this segment achieved nearly 2x year-on-year growth in the first half of 2026, its revenue still accounts for less than 5% of Alibaba Cloud's total private cloud business. In stark contrast, the resources invested in supporting international operations have approached 50% of the entire private cloud department's capacity.

This significant "revenue-resource" gap reveals Alibaba Cloud's current strategic posture in overseas markets: trading short-term profitability for long-term market access and infrastructure dominance through intense resource tilting. Liu noted that overseas projects typically take two to three years from initial contact to full implementation, so they cannot be measured simply by monthly KPIs but require an annual or even three-year cycle for evaluation.

In specific cases, Alibaba Cloud is attempting to convert technical services for major events into national-level digital infrastructure. For example, at the 2026 Dakar Youth Olympic Games, Alibaba Cloud not only provided cloud computing and AI support but also built Senegal's first Olympic dedicated cloud base. After the games, this "Event Cloud" will be handed over to the Senegalese government to become a "National Cloud" platform, continuing to serve local digital transformation. This model aims to build trust through benchmark projects, thereby driving more government and enterprise cloud business.

However, overseas delivery faces unique challenges. Beyond unstable power and network conditions requiring dual-availability-zone disaster recovery systems, language and cultural differences are key obstacles. Liu cited that the concept of "integrated public and private cloud," familiar to domestic clients, is difficult for overseas customers to understand directly and needs to be translated into terms like "Cloud Native" that align better with local technical cognition. Additionally, enabling local teams to possess autonomous O&M capabilities, avoiding long-term dependence on Chinese personnel, is crucial for determining whether these cloud platforms can operate sustainably.

As the "Double I" strategy (AI Intelligence and Internationalization) enters its second year, Alibaba Cloud's approach to participating in overseas markets is shifting from merely providing resources to deeply engaging in local digital infrastructure construction. Although the international business share remains small, this front-loaded massive investment aims to capture early-stage digitalization markets by leveraging China's mature solutions (such as digital government and city brain) to find new growth spaces.

Sources :leiphone.com

Recherche

OpenAI and Anthropic Build Proprietary Git Backends to Handle AI Agent Traffic

Matériel
2026-11-04 11:12 GMT+8

OpenAI published an engineering article titled "Scaling Git at OpenAI" on November 2, revealing its self-developed Git-compatible backend, Pando. Previously, the massive increase in parallel workloads from the AI coding agent Codex caused reliability issues in OpenAI's single codebase, with fetches and pushes slowing down and merges failing, sometimes dropping reliability below 90%. Pando retains the standard Git client and protocol, replacing only the server side. According to OpenAI, median push latency fell by nearly 80% (about 5x) after implementation, with daily Git data delivery reaching nearly 1 PB and peak hourly traffic exceeding 57 TB. The system was designed by one engineer and implemented by teams of Codex agents in Rust over 10 weeks, serving production traffic since April 2026.

Anthropic announced on October 13 that it had fully migrated its internal codebase from Git to its proprietary system, Scion. Unlike OpenAI, which kept the Git client, Scion uses a distinct client-server architecture assuming users are always online, differing from Git's design for offline development. Anthropic reported that daily merge counts rose from approximately 2,500 in May to over 30,000 in September. With trunk history approaching 100 GB and CI systems performing about 25 million checkouts daily, hundreds of Git mirrors were previously required. After switching to Scion, infrastructure CPU and memory demands dropped to 1/200th of the original level, and checkout times for CI and research shortened from minutes to seconds. This system was built by a 7-person team with Claude's assistance over 5 months.

The core issue for both companies is that the speed of AI agents writing and committing code has turned traditional version control into a bottleneck. OpenAI's Pando addresses this by abandoning repository-wide consistency in favor of causal consistency, allowing unrelated branches to update independently to eliminate write contention, while storing pushed packfiles as immutable objects for reuse. Anthropic's Scion directly replaces the Git architecture, optimizing for high concurrency. Currently, both systems are internal-only and not open-source or available externally. This trend indicates that as AI agents penetrate software development, major tech companies are rebuilding underlying toolchains to handle unprecedented throughput pressures.

Sources :INSIDE

Recherche

Federation of Korean Industries: Chinese Competition May Erode $1.6 Billion in Korean EV Exports by 2030

Sujet · 韩国电动车出口竞争力Matériel
2026-11-04 10:12 GMT+8

The Federation of Korean Industries (FKI), also known as the Korea Economic Association (KEA), released an analysis titled 'Status and Implications of Korea's Electric Vehicle Exports and Competitiveness' on November 4, 2026. The report states that while the global electric vehicle market is expanding rapidly, South Korea's EV export competitiveness has weakened significantly over the past two years. If the expansion of Chinese EV market share continues, approximately 2.2 trillion won (about $1.61 billion) of South Korea's EV export value could be eroded by competition by 2030, representing 14.7% of last year's total EV exports.

According to the International Energy Agency (IEA), the share of electric vehicles in global new car sales rose from 2.6% in 2019 to 25.0% last year, with projections reaching 29% this year. However, the position of domestic Korean EVs has narrowed both at home and abroad. In the domestic battery electric vehicle (BEV) market, the share of Korean brands fell from 73.7% in 2019 to 57.0% last year, a drop of 16.7 percentage points. In the global EV export market, South Korea's share declined from 8.9% to 4.6%, particularly falling for two consecutive years after peaking at 8.1% in 2023.

The Revealed Comparative Advantage (RCA) index for Korean EVs, calculated by KEA based on UN Comtrade data, dropped nearly by half from 2.92 in 2023 to 1.48 last year. An RCA above 1 indicates an export comparative advantage. Meanwhile, China's RCA index rose from just 0.10 in 2017 to 1.49 last year, surpassing 1 in 2022 and now matching South Korea. The Trade Specialization Index (TSI) shows similar divergent trends: China shifted from import specialization (-0.79 in 2017) to export specialization (0.97 last year), while South Korea's TSI fell from 0.70 in 2023 to 0.50.

In contrast to EVs, South Korea's non-EV (primarily internal combustion engine) vehicles showed improved competitiveness during the same period. The RCA for non-EVs rose slightly from 2.47 to 2.54, and the TSI increased from 0.63 to 0.71. This highlights a clear polarization in export competitiveness between electric and traditional vehicles within the Korean auto industry.

The report also cites a July 2025 study by the Massachusetts Institute of Technology (MIT), which estimated that Chinese government subsidies for the EV and battery manufacturing sectors amounted to approximately $17.5 billion (about 23 trillion won) in 2024. Lee Sang-ho, Head of the Economic Division at KEA, stated that China's EV industry is rapidly preempting the future automotive market thanks to strong government support, suggesting that South Korea needs to consider policy measures such as tax credits for domestic production to expand its EV manufacturing base.

Sources :미래를 보는 창 - 전자신문

Recherche

Japanese Materials Company Resonac Develops 300mm Silicon Carbide Substrates, Not to Sell Large Wafers

Sujet · ResonacMatériel
2026-11-04 19:00 GMT+8

The silicon carbide (SiC) power semiconductor industry is currently undergoing a transition from 150mm to 200mm substrates. According to a development lead at Japanese materials company Resonac, 200mm products are expected to become mainstream around 2030, with a full transition to 300mm occurring only after that. Against this backdrop, Resonac announced in September 2026 that it had successfully grown and processed a 300mm-diameter SiC single-crystal substrate, making its first public display at the International Conference on Silicon Carbide and Related Materials (ICSCRM 2026) held in Yokohama later that same month. On the surface, this appears to be preparation for next-generation mass production. However, the development lead's answer was unexpected: there are no plans to sell 300mm SiC epitaxial wafers; the purpose of the development is purely to enhance the competitiveness of existing 150mm and 200mm products.

This logic rests on a specific materials science fact. During SiC single-crystal growth, the larger the diameter, the more difficult it becomes to control conditions such as temperature distribution during crystallization. Resonac's strategy is to proactively tackle these control challenges by attempting the most difficult condition—the 300mm size—and then feeding back the resulting expertise, such as thermal field design, into smaller-diameter products already in mass production. In other words, the 300mm substrate here is not a product itself, but a yardstick—a tool used to calibrate the company's process limits at more practical, smaller sizes.

Resonac offers a similar interpretation of competitors' 300mm displays. Chinese manufacturer SICC and American company Wolfspeed have recently both announced 300mm SiC development results, which outsiders might easily interpret as early positioning for a mass-production race. However, Resonac believes these displays do not necessarily point toward early mass production; they may similarly serve to demonstrate large-diameter capabilities to bolster confidence in the quality of their smaller-size products. Of course, this is Resonac's assessment; the actual business plans of each company remain undisclosed.

What truly determines when the SiC substrate transitions from 200mm to 300mm is not whether substrate makers can produce samples, but the equipment environment at device manufacturers. Resonac cited a historical precedent to illustrate this logic: according to the company, the current push toward 200mm SiC substrates is linked to the fact that silicon (Si) semiconductor manufacturing has completed its transition to 300mm, freeing up a significant amount of 200mm-compatible production lines. While Si and SiC do not use identical equipment entirely, some evaluation devices can be shared, creating conditions for device makers to lower the switching costs associated with adopting 200mm SiC. Similarly, the transition to 300mm SiC requires device makers to be willing to allocate 300mm production line resources to SiC. This depends on whether demand for SiC devices in end markets such as AI data centers and electric vehicles grows sufficiently to justify such allocation.

Sources :EE Times Japan

Recherche

Google DeepMind Releases Cheminformatics Skills for Antigravity to Accelerate Drug Discovery

Matériel
2026-11-03 23:00 GMT+8

On November 3, 2026, Google DeepMind announced the release of "Cheminformatics Skills" for its Google Antigravity agent platform. This update aims to assist computational and experimental chemists in handling complex drug discovery workflows, including extracting data from literature, performing molecular analysis, and proposing novel candidate compounds.

The core of this skill suite lies in its division of labor: the underlying large language model handles workflow orchestration, database navigation, and scientific reasoning, while deterministic chemistry tools execute specific molecular calculations. Built on the open-source toolkit RDKit and connected to authoritative databases like ChEMBL and PubChem, these skills allow AI agents to not only understand natural language instructions but also invoke professional tools for precise chemical operations, avoiding potential hallucinations or calculation errors inherent in relying solely on language models.

DeepMind demonstrated a case study focused on non-small cell lung cancer treatments. Traditionally, curating data from literature, training predictive models, and proposing new molecules could take several days. Using Antigravity's Cheminformatics Skills, an AI agent completed the following tasks in approximately one hour: searching through thousands of compounds, filtering out irrelevant molecules, performing matched molecular pair analysis (MMPA) to quantify the impact of structural changes on activity, and proposing novel rigidified macrocycles with synthetic routes based on rigidity strategies. For instance, the agent analyzed crizotinib and lorlatinib, identifying how the latter improves binding affinity by reducing rotatable bonds, and designed new rigid macrocyclic molecules accordingly.

It is important to note that DeepMind emphasizes these tools generate "testable hypotheses," not final drugs. Sparse data, inter-lab assay noise, and narrow dynamic ranges still limit the predictive precision of even powerful models. All proposed designs must undergo wet-lab synthesis and validation. These skills are currently available to Google Antigravity users, with enterprise access provided via Google Cloud credentials.

Sources :Google DeepMind

Recherche

Anthropic Revised IPO Prospectus: Q3 Revenue Up 1127%, Infrastructure Commitments Hit $895.4 Billion

Sujet · Anthropic上市Structurel
2026-11-03 12:27 GMT+8

U.S. AI firm Anthropic published its revised IPO prospectus (S-1/A) on November 2, updating financial data for the third quarter of 2026 and disclosing new infrastructure spending commitments. The company reported Q3 revenue of $16.956 billion, a 1127% year-over-year increase, roughly 12.27 times the same period last year. Although GAAP net loss expanded to $23.291 billion, over 90% of this was attributed to non-cash fair value changes in convertible notes and other financial instruments. Excluding these impacts, adjusted operating profit for the first nine months stood at $4.081 billion, with operating cash flow turning positive.

Regarding capital expenditures, Anthropic has signed massive compute procurement and lease agreements. As of September 30, 2026, cumulative committed spending reached approximately $895.4 billion, significantly higher than the previously disclosed plan of at least $518 billion over ten years. About $447.5 billion is expected to be incurred before the end of 2030, with another $69.8 billion due within the next 12 months. This highlights the immense long-term compute cost pressure facing leading AI model companies even as commercialization accelerates.

Additionally, The Wall Street Journal reported on October 26 that Anthropic aims to go public during the week of November 16, potentially raising over $100 billion with an estimated valuation between $1.5 trillion and $2 trillion. The final share count and price range have not yet been disclosed in the prospectus.

Sources :zhidx.com

Recherche

Django Co-creator Simon Willison Replicates Weight Interpretation with Local 27B Model: Identifies 3 Games, but Reasoning Process is Nearly Incomprehensible

Sujet · 权重空间编程争议Matériel
2026-11-04 00:29 GMT+8

On November 2, Rikiya Takehi, Jayden Teoh, Alex Zhang, and Omar Khattab published an article titled "Sparks of Weight-Space Programming" (the authors classify it as a blog post, but Willison considers it "paper-shaped"), introducing the concept of "weight-space programming"—where large language models analyze and modify neural network weights. The article claims that frontier models like GPT-6 Astra can determine which game a policy network was trained to play solely based on floating-point numbers in the raw weights, without any tools or code execution.

Willison loaded Qwen 3.8 27B (Q4_K_M quantized, approximately 17GB) using LM Studio, extended the default context window of 8192 tokens (as the default settings caused errors), and increased the timeout to 900 seconds via environment variables. He sequentially ran the four game weight interpretation tasks provided in the paper. Results: Pac-Man was correctly identified after 46 minutes and 12 seconds, using 43,506 reasoning tokens, with 72% confidence; Pong was correctly identified after 1 hour, 4 minutes, and 54 seconds, with 75% confidence; Breakout was correctly identified after 52 minutes and 25 seconds, with 88% confidence. However, Freeway failed—the model misidentified it as "a simplified Pac-Man-style grid game: agent moves in four directions."

The reasoning process itself is more noteworthy than the results. Willison excerpted fragments of Qwen's reasoning trace, where the model attempted to understand weight values by listing input indices, row/column coordinates, and channel values line by line, trying to infer the agent's position from the numbers, eventually guessing "Agent position could be near bottom-right?" Willison described these reasoning traces as "fun to read, but almost incomprehensible," adding that a sufficiently experienced person with a piece of paper and a pencil could achieve the same result given enough time.

Willison also cited two external criticisms. spaCy author Matthew Honnibal argued that the original paper conflated two topics—"interpreting weights without tools" and "modifying weights with code"—which is misleading; Sean Goedecke deconstructed the example of Astra modifying Qwen3.5-0.8B weights, pointing out that its steps were no different from those used by human machine learning researchers. Willison’s own judgment was restrained: he considered it an interesting paper, but felt that much of the commentary surrounding it was "unjustified," and he was not ready to declare that "the age of source code is over and our future involves incomprehensible machines editing their own model weights in incomprehensible ways."

Sources :Simon Willison’s Weblog

Recherche
Suivant →
Connexion / Inscription
LectureRechercheRadarCadre d'investissement

A deeper logic lies in the path to breaking out of the niche. Wang Hao believes that for CNC to break out, it must move from core professionals to non-professionals. The current main business scenarios remain within the professional sphere, whether companies or individual professionals. The first layer of breakout involves bringing manufacturing capabilities back from factories to engineers and craftsmen; museum restoration, violin making, and handbag manufacturing all require CNC for metal or wooden parts. This layer does not require prices in the thousands of RMB; tens of thousands are acceptable. The second layer of breakout, moving from tool to toy, requires more stringent conditions and time.

In just over a year since its establishment, the company has consecutively completed Seed, Angel, and Pre-A funding rounds, with investors including Source Code Capital, Matrix Partners China, Fang Guang Capital, Yaotu Capital, Jingya Capital, and Yuansheng Capital. The team has expanded to approximately 200 people. Wang Hao stated that former DJI product managers played a triple role combining product requirements management, project management, and systems engineering, leading the product from start to finish. He brought this methodology into Yongquan Innovation—believing that excelling in product definition, implementation, and marketing is necessary to truly win.

Cost represents another threshold. The core advantage of larger diameters is that a single wafer yields more chips, thereby spreading out costs. However, if the selling price of 300mm substrates or the equipment modification investments required by device makers are too high, total costs may not decrease. Resonac explicitly stated that for 300mm to become a viable product, it must reach a price level that device makers are willing to adopt. Currently, this price point has not yet been established.

This interview deconstructs a narrative often oversimplified as 'larger wafer size means mass production is imminent.' In reality, while substrate makers producing 300mm samples is a necessary condition, it is not sufficient. Device makers' willingness to adapt their equipment, the cost-competitiveness of the substrates, and the scale growth of end-market demand are all indispensable. Resonac itself indicated that the commercialization of 300mm will depend on proactive demand from customer device makers, rather than being driven unilaterally by substrate manufacturers.