OpenCode Responds to Model Weakening Controversy; Microsoft Open-Sources 4B Desktop Agent

AI news is always dense, but what's truly worth paying attention to is often not the release itself, but the controversies, shifts, and hidden risks behind it. Today several stories point in different directions: model capability disputes, agent deployment, cautious sentiment in capital markets, and new issues in supply chain security.

Model Capability Controversy: OpenCode Says DeepSeek Wasn't Weakened

There's now a response to the controversy surrounding DeepSeek^Ad V4.1 Flash. OpenCode developer Dex stated that the model itself wasn't weakened—rather, OpenCode's inference efficiency is so high that many users mistakenly believe capabilities have shrunk. The team emphasized that only the execution approach differs from DeepSeek's official one, and they will soon provide evidence to prove it, while inviting users to report specific poor experiences to help with troubleshooting. OpenCode had just announced $100 in usage credits for the model.

This kind of controversy reflects a common issue: the same model can perform noticeably differently under different inference frameworks. The "dumbing down" users perceive may not come from weight changes, but from differences in inference paths, context management, or tool-calling strategies. For developers relying on third-party frameworks, establishing a reproducible evaluation baseline is more reliable than judging from a single experience.

Microsoft Open-Sources 4B Desktop Agent: Small Models Can Get Work Done Too

Microsoft has open-sourced a 4B AI agent called Mage (referred to in the source material as MajorVL Co-Work), which scored 66% on OSWorld Verified—close to the human baseline of 72.4%—and outperformed a model nearly 6x its size (63.9%). Its key capability is autonomously deciding when to click through the interface like a human and when to directly type terminal commands, handling complex desktop tasks with a single policy. It's 9.5GB in size and can run on a Mac mini.

What's noteworthy about this direction is "deciding which interaction method to use." The hard part of desktop agents has never been single-step operations, but choosing the right execution channel when facing a task. A small model achieving near-human performance means localized, low-cost desktop automation is becoming feasible.

OSWorld is a benchmark environment for evaluating computer-operation agents. It requires completing a series of multi-step tasks in real Ubuntu, Windows, and macOS desktop systems—such as editing spreadsheets, configuring system settings, and copying information across applications. Unlike benchmarks that only test single-step clicks or simple Q&A, OSWorld emphasizes long-horizon planning and real interaction. Task success requires agents to understand screen content, choose input methods, and handle unexpected pop-ups. Its human baseline is about 72.4%, and approaching that level means the model already has considerable usability in real desktop scenarios. This benchmark is often used to compare the actual deployment capabilities of different desktop agents.

Capital and Organizations: Funding Cools, Talent Flows Accelerate

OpenAI's new funding round target has been lowered. According to The Wall Street Journal and Financial Times, it went from seeking $30 billion at a valuation of about $1.4 trillion less than a month ago, down to about $20 billion at a valuation of $1.25 trillion. Some large investors showed lukewarm interest, and reports say about $11 billion has been secured.

Anthropic has been more active: it filed IPO prospectus documents and plans to invest nearly $900 billion in debt in areas such as server buildings and power in the future; its balance sheet shows $117.5 billion in liabilities and $112 billion in cash, with off-balance-sheet obligations such as lease contracts and data center power and GPU procurement agreements not directly reflected. It also acquired Mithril Medicine, a biotech company with only one full-time employee. Founder Cody Tranbarger had previously run the anonymous X account Professor Oak and went public this week. Mithril is developing a drug for a rare childhood malnutrition disorder that currently has no approved therapy, coordinating 11 external labs through the cloud and achieving early positive results in mouse experiments in under 18 months.

On the talent front, ByteDance Seed core member Zhou Chang has left to start a company in Physical AI and has secured backing from a top-tier USD fund. Zhou Chang holds a bachelor's degree from Fudan University and a PhD from Peking University, previously served as technical lead for Alibaba's Tongyi Qianwen, and joined ByteDance Seed in 2024 to oversee multimodal and world models, taking charge of visual generation such as SeaDream and SeaDance as well as robotics. In the second half of the year his scope of management shrank, with some engineering work transferred to Zeng Yan.

Products and Hardware: Grok Adds Alipay, ChatGPT Supports Multiple Accounts

Grok's subscription checkout page now offers Alipay payment, adding an option beyond bank cards and Apple Pay. The Premium plan is 50% off for the first month—originally $30 per month, now just $15 for the first month, returning to the original price the following month. The plan now bundles the benefits of Grok, Cursor, and X Premium.

ChatGPT officially supports multi-account switching, currently rolled out on mobile with desktop to follow. Users can log into up to two accounts simultaneously without repeatedly logging out, with chat history, memory, files, and subscriptions remaining independent. Teebo also teased that reset cards will be given out over the next two days.

OpenAI's API pricing page now shows GPT-6 Astralog: $10 per million input tokens, $50 for output, and $1 for cached input, with the discounted pricing lasting at least until December 31. It's the previously announced legal-specialized version, equipped with a dedicated legal retrieval index and analysis tools that can assist with case law research and contract review. Currently the API is only open to select law firms, and the official launch date has not been confirmed.

Security and Governance: Pretraining Backdoors and New Regulatory Developments

Anthropic, together with UC Berkeley, UIUC, and other institutions, published a paper that for the first time end-to-end demonstrates that poisoning only the pretraining corpus can implant a remotely triggerable backdoor into an agent model. In the experiment, poisoning about 0.1% of pretraining data resulted in an agent backdoor attack success rate of about 92% on a 4B-parameter model. This means the pretraining data supply chain itself can be tampered with, and the trigger can be stored externally to the model long-term.

Another questioned story: someone on a dark web forum claimed to have obtained the weights of Anthropic's unreleased model Claude Fable 5—about 8TB for the full model, split into 640 shards—and began selling them. The poster claimed to have downloaded about 7GB of one shard from a misconfigured AWS S3 bucket associated with Amazon Bedrock, after which the bucket was restricted. But the technical community quickly questioned this as more likely a fake scam, and the credibility of the so-called leaked files is doubtful.

On the policy front, Google announced that starting in 2027 it will voluntarily adopt the AI training control proposal being developed by the IETF AI-PREF working group, giving website publishers and creators a unified, simple control mechanism. Websites can declare through Robots-related protocols whether their content may be used for AI training; once implemented, Google's approach will align with that of Bing, OpenAI, Anthropic, and others. A draft bill in Delaware would allow AI agents to autonomously operate companies, sign contracts, and file lawsuits, with company owners not liable for damages caused by agents unless fraud or insufficient capital contribution is involved. The Secretary of State called this a boon for state finances, since business registration fees account for about a quarter of the state budget. The House Oversight Chairman has requested a review of the plan.

France's National Assembly passed an amendment to the 2027 budget bill by a vote of 11 to 10, preparing to charge content service providers transmitting data to French users 0.055 euros per GB, with an official target of collecting up to about 10 billion euros a year. Netflix, Google, Meta, Amazon, YouTube, and cloud service providers are all potentially in scope, and services like ChatGPT and Cloud as well as the many agents that transmit images and videos could also be included.

Pretraining backdoors differ from common fine-tuning poisoning: the attack occurs in the early stage when the model learns general language capabilities, and once polluted data is mixed into massive corpora it is extremely difficult to audit line by line. Because the backdoor pattern is encoded into parameters during pretraining, subsequent safety alignment and fine-tuning often cannot completely remove it, and trigger conditions can be designed as specific external inputs or context patterns that remain completely inactive normally. The stealthiness of such attacks lies in the fact that the model performs normally in routine evaluations and only executes malicious behavior under attacker-defined trigger conditions. Defense challenges include training data provenance tracking, corpus deduplication and anomaly detection, and behavior monitoring during deployment.

Quantum Computing and Storage: Don't Get Misled by Headlines

Google's quantum chip completed in 5 minutes a task that would take a classical supercomputer 10 to the 25th power operations—the result of a random circuit sampling experiment. Some physicists say it supports the many-worlds interpretation, but saying it "proves many-worlds" is media exaggeration of the original conclusion; the original conclusion is only that quantum computing far exceeds classical computers on a specific task.

On storage, financial reports disclose that DRAM products on the sixth-generation G6 technology platform completed engineering sample validation in Q3. According to leaks, the previous-generation G5 entered mass production in September and is expected to reach mass production within a year, narrowing the process gap with overseas giants. The report also shows Q3 revenue of about 118.7 billion yuan, up 19.3% quarter-over-quarter; net profit of about 89 billion yuan; gross margin of about 89.5%; and R&D investment of about 11.07 billion yuan in the first three quarters, up 73% year-over-year. As for whether the progress comes from shifting to a 4F structure or more multiple patterning, that is currently only speculation.

Taken together, today's theme is not any single breakthrough, but capability disputes, local agents, cooling capital, and data supply chain risks occurring simultaneously. For developers and content creators, understanding the mechanisms behind these changes is more valuable than chasing headlines.

Random circuit sampling is a common class of benchmark tasks in quantum computing, designed to have a quantum processor execute random quantum circuits that classical computers cannot efficiently simulate, then sample the output distribution. This is precisely the task Google used earlier when it claimed "quantum supremacy" on the Sycamore processor. What needs to be distinguished is that the experiment proves a quantum device is faster than a classical supercomputer on one specific computational task—not that quantum computers have general-purpose computational advantages, and certainly not anything about verifying physical interpretations. The many-worlds interpretation belongs to the philosophical and physical interpretation of foundational quantum mechanics issues and has no direct proving relationship with the results of random circuit sampling experiments.

One-line Summary

Model capability disputes, local desktop agents, cooling funding, and pretraining backdoor risks all surfaced on the same day

Paragraph Summary

The article rounds up several notable developments in AI today: OpenCode responded to the DeepSeek V4.1 Flash experience controversy, saying it's a misjudgment caused by inference efficiency differences rather than model weakening; Microsoft open-sourced the 4B desktop agent Mage, which reaches 66% on OSWorld Verified, close to the human 72.4%; OpenAI's funding target was cut from $30 billion to $20 billion with valuation dropping to $1.25 trillion, while Anthropic filed for IPO and borrowed heavily to invest in infrastructure; Anthropic and universities for the first time end-to-end proved that poisoning only about 0.1% of pretraining corpus can implant a remote backdoor, with a ~92% attack success rate on a 4B model; other items include Grok adding Alipay, ChatGPT multi-account support, Google's AI training control proposal, France's per-GB fee amendment, and clarification of quantum computing headlines.