Musk Sabotages Ukraine and China Restricts Open-Source Models

Musk apparently hatched the plan to sabotage Ukraine

On October 14, Ukraine lost access to essential parts of the Starshield satellite service without warning. The connection dropped as soon as drones or vehicles were traveling faster than 100 miles per hour, in some cases as low as around 70 kilometers per hour. This primarily affected the interceptor drones used by Kyiv to intercept the jet-powered Shahed and Geran models, as well as attacks on Russian supply lines between 30 and 200 kilometers behind the front. According to their own reports, Ukrainian units destroy more than 700 supply trucks there per month. A Ukrainian officer compared the outage to a nuclear strike in comments to reporters, saying all operations had come to a standstill. Zelensky said in Brussels on Friday that they were unable to intercept a number of incoming Shaheds, and that the supply of water and food to troops was also disrupted.

The question of who ordered the Throttling has since been the subject of conflicting reports. On October 14, Musk wrote on X that Starshield was “owned and controlled by the U.S. government,” and that limits on speed or usage were decided by the U.S. military and ultimately the President. Two U.S. officials and one Western representative stated that the directive came from the White House, and that the Department of Defense had received the message that the service for the Ukrainian military must be slowed down. The White House stated that no such request had come from the President, but left open whether someone else had made it. Four people familiar with the discussions report that Musk himself proposed the speed limit to Trump, and Trump agreed. After a wave of attacks in Kyiv described as a “night of carnage,” Musk reportedly feared negative consumer reactions in Europe and publicly shifted the blame to the White House. On October 16, Trump said he didn’t know, it wasn’t him, and he didn’t think it was Musk either.

The outage coincided with an escalation between Washington and Kyiv. Trump called the Ukrainian attacks on Russian oil refineries “a disgrace” and blamed them for rising oil prices. Energy analysts, however, point to the U.S.-Israeli war against Iran as the main driver for diesel exports. The White House also threatened to halt intelligence sharing if Ukraine did not stop the refinery attacks. In the same week, Musk publicly called for Zelensky’s resignation. Nine senators from both parties, including Jeanne Shaheen, Thom Tillis, and Richard Durbin, have asked the Pentagon’s Inspector General for an investigation: American embassy and security personnel also reportedly stopped receiving warning messages during Russian air raids.

Zelensky stated on Friday that the original speed has not yet been restored, but the process is underway. During his meeting with EU heads of state and government, he asked for an additional 70 billion euros for 2027, estimating the war costs for the coming year at 155 billion euros. Commission President von der Leyen confirmed the 70 billion euro gap but expects other partners to contribute a portion. Currently, payments are flowing from a 90-billion-euro package approved in December; due to renewed financial pressure, the 210 billion euros in Russian assets frozen in the EU are once again up for debate. → The Atlantic, Gizmodo, Diplomatic,, The Guardian, The Kyiv Independent

Synthszr Take: The interception rate against incoming Shaheds depended on a speed limit that could be set overnight in Washington without anyone in Kyiv being forewarned. This is the price of a dependency that was never specified as such in the contract: a service that can be throttled via a configuration change without a single device having to be returned. Europe is financing a 90-billion-euro package whose most effective component can be throttled by a single private provider with a parameter from 160 to 70 kilometers per hour. The consequence is inconvenient and expensive: a second constellation, a second operator, and the ability to switch over in the field within minutes, even if it doubles the cost. October 14, 2026, marks the date in the procurement files of European defense ministries when sovereignty shifted from a conference topic to a budget item.

China wants to ban the export of large open-source models

On Friday, China released a draft proposal that would ban or require approval for the export of 40 technologies in the fields of artificial intelligence, chips, solar technology, and biomedicine. It was jointly issued by the Ministry of Commerce and the Ministry of Science and Technology as an update to the national catalog of export-restricted technologies. 14 technologies are to be completely banned, with another 26 subject to approval. A one-month public consultation period has been running since Friday. The Ministry of Commerce called the move necessary, citing national security as one of the reasons.

23 of the listed technologies concern artificial intelligence, from advanced large models to agents. The export of core technologies for large models trained on sensitive government data is to be banned, as are agents for highly sensitive applications such as defense and critical information infrastructure. Models with a parameter count of two trillion or more, or with certain world-leading capabilities, could be restricted. Chinese models like Moonshot’s Kimi K3 and Alibaba’s Qwen 3.8-Max are above this threshold, although the draft does not name any specific models. In addition, several chip design technologies are included, such as those for AI processors and memory chips.

Twelve items relate to the solar industry, including processes for producing polysilicon, the key raw material in the solar supply chain, and technology for certain solar cell modules. Five other entries concern biomedicine: platforms for cell and gene therapies, processes related to small interfering RNA for silencing disease-causing genes, and technologies related to pathogens. The draft targets the transfer of ownership of these technologies but leaves licensing agreements, collaborations, and outsourced research work unaffected. Such licensing deals have fueled China’s rise as a pharmaceutical nation, with AstraZeneca and Pfizer among the corporations that have licensed drug candidates from Chinese developers. According to government data, Chinese drug developers signed 81 licensing deals worth up to $110 billion in the first half of 2026.

The move comes against the backdrop of years of U.S.-led restrictions that block China’s access to the most advanced AI chips and the machinery to produce them. In parallel, both sides had recently reported progress in their dialogue on AI safety and a reporting mechanism for security-related AI incidents. In June, the U.S. Commerce Department had requested Anthropic to block foreign nationals' access to the Fable 5 and Mythos 5 models; the requirement was later lifted. An analyst at the research firm Omdia classifies the Chinese draft as a response to Western sovereignty demands in AI, robotics, and telecommunications. An economist at the bank Natixis points to increasing fragmentation not only in trade and investment but also in access to technology. → Reuters, Associated Press, Quartz

Synthszr Take: A limit of two trillion parameters is a customs barrier drawn right through a weights file. Beijing is adopting the same method it has suffered under for years and extending it upwards: Washington blocks lithography machines and chips, while China blocks polysilicon and the models themselves. The fact that licensing deals remain exempt (81 in the first half of the year, with a volume of up to $110 billion) reveals the calculation behind it: foreigners can continue to buy the results, but the ability to produce them stays in the country. European companies that currently base their products on freely available Chinese weights are thus factoring in a supply chain that could become subject to approval after the one-month consultation period expires. Model weights are now an export commodity like polysilicon, and supplier risk must therefore be part of every architectural decision made this week.

Qwen overtakes DeepSeek in China

Alibaba’s assistant app Qwen has replaced DeepSeek as China’s second-largest AI chatbot. According to the QuestMobile 2026 report on AI users and marketing, picked up by 36Kr, in August 2026, 22.8 percent of all users of AI-native apps opened DeepSeek at least once a month, down from 50.1 percent in August 2025. Qwen, formerly known as Tongyi, climbed from 1.1 to 29.4 percent in the same period. The leader remains Doubao from ByteDance, with an increase from 54.9 to 77.2 percent. QuestMobile calls this metric Active Penetration, which is the percentage of all users of AI-native apps who opened a specific app within the month. → Hello China Tech

Synthszr Take: DeepSeek has lost around 22 million users in twelve months, while the total market has nearly doubled from 287 to 537 million, and the model has not gotten any worse during this time. Qwen and Doubao are sitting on distribution machines: Alibaba with its e-commerce and browser touchpoints, ByteDance with Douyin, plus paid installations on a scale that a research lab like DeepSeek simply isn’t willing to match. Technical superiority got DeepSeek global press coverage in January 2025, but the press doesn’t install apps.

Beam: Another open-source alternative to Chinese models

Reflection AI has released the weights of its first open model, Beam, positioning it against the Chinese open models that many US companies are turning to due to lower costs. Beam has 501 billion parameters, of which 23 billion are active parameters per token, and according to the company, it targets coding and agent tasks. According to the provider, the selling point is not the token price, but the number of tokens the model needs to complete a task. To achieve this, Reflection has implemented a penalty for superfluous tokens in its Reinforcement-Learning training; CTO Ioannis Antonoglou told VentureBeat that a naively set penalty simply teaches a model to no longer solve the task. According to their own statements, one of the largest RL runs by an Open-Weight lab took place: over 100 million rollouts in four weeks on a cluster of 10,500 Nvidia GB300 GPUs, plus around 1.3 billion sandboxes and a peak of 883,006 virtual CPUs across 23 clusters on Google Cloud and AWS. → VentureBeat

Synthszr Take: The calculation that hardly anyone in purchasing makes: price per million tokens times tokens per task, and only the first factor is on the price sheet. Antonoglou identifies the method’s weakness himself: a naively set token penalty trains a model not to solve the task at all, because brevity is easier than doing the work. Reflection needed over 100 million rollouts in four weeks on 10,500 GB300 cards to achieve both conciseness and solution quality, which gives the word efficiency a very expensive backstory here.

Trump’s “Super Intelligence Force” meets for the first time to secure US lead

The Trump administration has established a new working group called the “Super Intelligence Force,” which, according to RedState, has convened for its first meeting. The panel’s mission is described in the report as a balancing act: to contain the development of artificial intelligence without damaging the USA’s technological lead. The article by Teri Christoph is dated October 16, 2026. The term superintelligence refers to hypothetical systems that surpass human capabilities in almost all areas; a binding, regulatorily useful definition does not yet exist. The report does not mention any concrete regulatory proposals, deadlines, or a list of members.

Synthszr Take: A committee named the “Super Intelligence Force” already has its problem in its name: it’s supposed to contain something that no one has clearly defined yet, while simultaneously not touching anything that generates revenue today. In practice, regulation takes hold where it can attach to something measurable, like computing power and chip exports. In early September, the EU showed that oversight can also attach to products when ChatGPT came under stricter DSA scrutiny, and no model slowed down as a result.

Meta’s Muse Spark Open requires real name and date of birth before download

Meta has released the Muse Spark Open model on Hugging Face with 816 billion parameters, of which 42 billion are active per query. The model card lists the release date as October 15, 2026. The accompanying Muse Model License permits use, copying, modification, fine-tuning, distillation, and the sale of copies, subject to two conditions: The interface of any product based on the model or its derivatives must prominently display “Built with Muse Spark Open.” Anyone offering the model as a Model as a Service via an API and generating more than $20 million in gross revenue with their affiliated companies over twelve months must first enter into a separate written agreement with Meta; internal use, research, and proprietary product features are exempt. Meta explicitly claims no ownership of the generated outputs. → Hugging Face

Synthszr Take: The download starts with a form that asks for more information than some loan applications: legal name, date of birth, organization name with all company identifiers, abbreviations prohibited. You can’t correct anything; the form is frozen after submission, and a typo in the company name could cost you access to Muse Spark Open and other models on the platform. The weights are free, but the path to them leads through an identity check, the result of which ends up in Meta’s data processing system and, according to the notice, may also be shared.

Meta’s new terms of service will allow scraping of news sites from October 30

Meta has revised its commercial terms of use: Starting October 30, owners of business accounts will allow the company to crawl websites associated with their products or services. The clause explicitly mentions web scrapers, bots, spiders, and crawlers and allows Meta to index, cache, and use the collected content to train its own products; “AI at Meta” is listed as one of these products elsewhere in the terms of use. This affects virtually all news organizations, as nearly every newsroom maintains at least one business account for Facebook pages, Instagram profiles, ads, or WhatsApp channels, including Nieman Lab. Three legal experts reviewed the text at Nieman Lab’s request and consider the wording broad enough to cover training on publisher content. Matt Topic of Loevy & Loevy calls it remarkable to hide such a provision in the terms of use, as it concerns material that would cost many millions of dollars in a licensing market; whether a court would enforce the clause is an open question, but publishers should assume it will be. → Nieman Lab

Synthszr Take: On October 30, publishers will lose the only currency they still had with Meta: the ability to say no and name a price. Topic values the same material in the licensing market at many millions of dollars, and this market is being closed here by a clause rather than settled with a check. The alternative offered is to no longer touch Facebook, Instagram, and WhatsApp for business purposes, meaning giving up distribution to protect content.

Parallax bundles Claude Code, Codex, and Cursor on the Mac, with human approval

Conifer has released Parallax, a Mac application that runs multiple coding agents simultaneously in a single interface, including Claude Code, Codex, Cursor Agent, Gemini CLI, OpenCode, and pi. A central element is a control instance called Captain: It breaks down a request into subtasks, distributes them to different agents, and retrieves a session if it deviates from the original task. In the provider’s example, three agents work in parallel on a CSV export, a webhook error, and the corresponding tests, ultimately delivering two Pull Requests. Before executing critical steps, the session pauses and waits for user approval. It also includes a review feature that, according to Conifer, checks not only the diff but the entire session history against the original instruction; a router selects verifier models from labs other than those that wrote the code. → Techpresso

Synthszr Take: The most striking element of this application is a button labeled 'Approve'. When Claude Code, Codex, and a Cursor agent are working on the same request simultaneously, the workday consists of approvals, cancellations, and queries, with very little typing involved. With 2.24 merge-blocking findings per pull request and three ongoing sessions, seven findings quickly pile up that someone has to read, classify, and take responsibility for.

AI has killed the six-page memo culture at Amazon

A former Amazon manager describes on his Substack 'A Life Engineered' that the company’s legendary writing culture has collapsed due to generative artificial intelligence. As evidence, he recounts a scene from a review meeting: An L7 Principal Product Manager began presenting a six-page document after 30 minutes, and within two minutes, it was clear to everyone in the room that he was reading the text for the first time himself. An L6 employee from his community was sitting in the room; no one dared to address it because the presenter was the highest-ranking person present. The author, who claims to have been with the company for almost twenty years, refers to Jeff Bezos’s justification for the 2004 PowerPoint ban: You can’t write a narratively structured six-page memo without thinking clearly. That guarantee is gone because six pages can now be generated without any prior thought. → Substack

Synthszr Take: The L6 in that room saw what everyone else saw after two minutes, and yet remained silent. That’s the real cost of this story: an hour of time from half a dozen expensive people, written off because hierarchy outweighs the obvious finding. In the past, the writing process itself prevented this situation; unprepared documents wouldn’t even make it into the room, and no one had to muster the courage to point it out.

Ex-OpenAI researcher declares AI neurons readable, splitting the safety research community

Former OpenAI researcher Nick Cammarata published a report on Thursday in which he claims that neural networks are far more understandable than the expert community previously assumed. “Almost the entire network is interpretable, if we just look closely enough,” he writes. Cammarata, who left OpenAI in 2023 and now works at the Interpretability-startup Goodfire AI, is thereby attacking a theory he himself co-founded.

At the center is a famous neuron from the image model InceptionV1, which responds to both cat faces and car fronts. Since the paper “Zoom In” by Cammarata and Chris Olah, this neuron was considered proof of superposition: the assumption that a network packs unrelated concepts into the same neuron when space is limited. Today, Cammarata reads the same neuron as a detector for V-shapes. Pointed shapes like fox snouts or flames activate it, while round shapes like the letter O push the activation into the negative. He explains the remaining ambiguity with a pattern he calls “recursive generalization”: A neuron for the letter sequence “iy” later responded to “diy” and finally to “do it yourself,” even though it no longer contains “iy”.

For language models, Cammarata provides preliminary figures. In Qwen3-32B, around 90 percent of neuron poles reportedly have a main function, about 7 percent have two related functions, and only 2 to 4 percent remain unexplained. For comparison: In his automated interpretability project at OpenAI in 2023, only about 1.7 percent of neurons were considered understandable, even by a generous standard. Goodfire claims to see similar results in models with up to trillions of parameters. The majority of the analysis was performed by AI agents that process neurons in large numbers. Cammarata does not present a fundamentally new method: He returns to the field’s original approach of directly examining neurons, just with better tools and larger models.

Reactions are divided. Until now, many labs have been training a second model, a Sparse Autoencoder, to separate the supposedly entangled concepts, and then studying this auxiliary model instead of the original. Anthropic’s Chris Olah has previously said that a large part of the neural universe could be effectively unobservable dark matter, and Cammarata himself once pessimistically ended his work on interpretability. In the safety debate, a lot hinges on this question: If networks were truly readable, deceptive or manipulated parts of a model could be detected and switched off during operation, instead of just observing behavior from the outside. → Hyperdimensional, Business Insider, The Decoder

Synthszr Take: An interpretability startup reports that interpretability is much easier than thought, and the industry is calling it a bombshell. The numbers warrant re-checking: 2 to 4 percent unexplained neuron poles in Qwen3-32B, collected mainly by AI agents labeling neurons in series, without peer review and without independent verification. Reinterpreting the cat-car neuron as a V-shape detector is methodologically plausible and elegant, but as a correction of a single 2020 finding, it doesn’t yet constitute a theory of how a 32-billion-parameter model thinks. The fact that the news “almost everything is readable” comes just as labs need to prove to regulators that they have their systems under control is a coincidence that sells well. If another lab reproduces the same 90 percent on a different model, it’s science; until then, it remains a well-founded hope with an attached business model.