From Models to Machines: DeepSeek’s Next Bet Is Physical AI

Feeding unpublished core ideas into ChatGPT could result in them being directly copied. Recently, Tristan Buckmaster, a mathematics professor at New York University, publicly warned against inputting original research into OpenAI tools. He and researchers from Anthropic spent over half a year tackling the $1 million Navier-Stokes problem, previously using a paid Codex account to run code, with all drafts stored on OpenAI’s servers. The professor revealed that before disclosing his real name, OpenAI had made no progress on the problem—but within a single day after he did, OpenAI’s model suddenly advanced to exactly their current stage, followed by the immediate publication of a mathematical proof. Although OpenAI insists prompts are never used to train its systems, the training switch on researcher accounts remains enabled by default, making this coincidence suspiciously precise.

Clude is building towards a world for specialised local models. This is only where DeAI can seriously compete against the trillion/billion dollar labs. 🫡

In the past few months, progress has slowed after the initial rapid advancements, as OpenAI and Anthropic have implemented stronger anti-distillation measures, increasing the cost and difficulty of distillation by tenfold. For example, OpenAI’s GPT-6 no longer directly exposes reasoning chains, forcing distillation efforts to rely on significantly more rounds of questioning to reconstruct the model’s thought process. Information that previously required only 2–5 interactions now often demands 10 or even 30 rounds, causing API invocation costs to rise by more than fivefold. Anthropic’s new defenses reduce reasoning tokens by 90–99%, including restrictions on transferring reasoning logs across conversations, models, and accounts; deploying a two-stage classifier to detect and block extraction attempts; and implementing rate limiting or account bans for repeated probing.

the end result was trash too. slopus did this in an hour and a half, way better. it's so unfathomably over for openai

Anthropic published a study predicting that small models may, within a few months, gain the ability to replicate autonomously without relying on data centers. They actually purchased 15 cheap Android TV boxes, each costing $55, and stitched them together to run a 26B model—at a speed of just 0.39 tokens per second, painfully slow. In another isolated experiment, GLM-5.2 has already demonstrated cross-device propagation within a simplified environment. While there remains a clear gap between this and full autonomous replication in the real world, Anthropic’s projections suggest small models could approach this capability threshold by the end of 2026. The entire world is racing to secure GPUs and spending hundreds of billions of dollars building data centers—yet AI might soon go parasitic on your home TV box. But upon closer reflection, the truly alarming implication is that such AI could fundamentally disrupt our familiar economics of compute. Today, whether it’s OpenAI or an AI startup, every generated token carries a cost: higher compute prices mean higher fees; interrupted funding means shutdowns. But a self-replicating malicious AI could theoretically offload its marginal operating costs indefinitely onto strangers. It doesn’t need users, doesn’t care about revenue, and doesn’t even need to buy servers itself. Find one vulnerable device, and it gains another free unit of computation. Infected devices pay its electricity bills, provide bandwidth, and absorb hardware wear. Yet the cost of cleaning up these infections falls on countless individual device owners, operators, and manufacturers. Traditional botnets have long stolen compute power—but AI introduces a new complication. Previously, hackers had to write custom exploits and deployment scripts for specific devices; in the future, models may autonomously study unfamiliar hardware, solve environmental challenges, and even pass on failed strategies to the next generation. The attacker’s expansion cost keeps falling, while society’s defense cost may rise ever higher. So I believe the most concerning takeaway from this research isn’t the grand narrative of AI runaway. If such entities ever emerge, we may for the first time face software that has no business model, no legitimate users, no identifiable operator—yet can persist indefinitely. They face no pressure to profit, and no economic incentive to shut down. After years of research aimed at making AI help us earn money, we may ultimately enable AI to achieve financial freedom first—using our money to do it.

Last year, Anthropic paid $1.5 billion in settlement over pirated books, and now it’s busy teaching the entire AI industry what “stealing knowledge” means. I just saw Anthropic announce further progress in its anti-distillation technology, claiming it has significantly reduced the effectiveness of distillation attacks against Claude over the past few months. Previously, it publicly named Chinese AI companies—including DeepSeek, Moonshot AI, Alibaba, and Zhipu—accusing them of training their own models by大规模 calling Claude. Now, Anthropic not only aims to identify and block suspicious traffic but is also restricting the extraction of reasoning processes at the model level. It even secretly degraded response quality for certain requests to deter distillation—only adjusting after public criticism. When training large models, internet articles, books, and code were all used as training material. AI companies argued that machines learning from human knowledge was both technically and legally justified. Yet today, when other models learn from Anthropic’s own responses, it suddenly becomes a critical security threat that must be fiercely guarded against. Anthropic is fully entitled to protect its business interests. What I find objectionable is its tendency to frame commercial competition as an AI safety issue, attempting to legitimize rules dictated by industry leaders. Moreover, I believe Anthropic’s highly visible anti-distillation efforts reveal a troubling truth: the commercial moats of top model companies may be far more fragile than they’re willing to admit. If leading intelligence can be replicated at low cost, then the massive training investments and pricing power of closed-source models will be hard to sustain indefinitely. An industry that grew by absorbing all of human knowledge is now seeing its most powerful companies attempt to prevent knowledge from continuing to spread. This is something I believe everyone should be deeply wary of.

I broke my back because I expect AI to cure paralysis

Lately, ChatGPT has been extremely slow, and my $200 Pro account seems to have been “dumbed down,” frequently hitting rate limits. Yesterday afternoon, I finally couldn’t take it anymore—I upgraded to a new Plus account and asked it to draw a pelican, and it suddenly worked much better. But when I tried to work again in the evening, I got another notification saying my quota was exhausted. Based on past experience, I wonder if a new model is about to be released. Recently, ChatGPT has been thoroughly outperformed by Claude’s Opus 5.5; many content creators are now generating entire videos directly with Opus 5.5. I have to admit, Claude’s aesthetic sense and coding ability are truly impressive. However, Claude accounts are notoriously prone to mysterious bans—I had an old Pro account from April 2023 that I’d been using for nearly three months, and then it was suddenly banned without warning. Here are my strategies for avoiding or recovering from account bans: Claude Account Ban Prevention: There are many anti-ban tips online, but I’ve been reflecting on why my account was fine at first and then suddenly banned. Before the ban, I had consistently hit Claude’s 5-hour usage limit. I didn’t use IP isolation or any physical anti-ban measures—so I suspect the issue was simply exceeding the 5-hour cap. Therefore, if you use Claude, I recommend always leaving some buffer in your usage. Once an account is banned, the chances of getting it reinstated are very low. I appealed my Pro account, but so far there’s been no response. ChatGPT Anti-Dumbing: ChatGPT doesn’t really ban accounts—it just “dumbs down” and throttles usage. I even contacted support, and they assured me my Pro account wasn’t flagged or downgraded. But my Pro account clearly struggled when asked to draw a pelican. A friend in the group suggested letting the account rest for a few days—and it worked. Maybe this is similar to Claude: don’t overuse it. My $200 ChatGPT Pro account is used across Windows and Mac, logged in via Feishu authorization across three different entities. So if you’re experiencing what I am—apparent dumbing-down or throttling—it’s likely your account has been placed into a restricted pool and put in “time-out.” Resting the account probably helps. If you rely on both Claude and ChatGPT for work like I do, build up backups. For example, maintain multiple Google accounts and stagger your use of free tiers for both ChatGPT and Claude. Let each free account mature for 2–3 months, and maintain 2–3 active free accounts at a time to create three tiers of accounts. Finally, I strongly recommend using both domestic and international models in parallel. International models excel in raw capability, while domestic models offer superior ecosystems. Both are worth becoming familiar with.