The hypocrisy of the US and the Innovation of China

Welcome back to the Avicenna AI newsletter.

Here’s the tea 🍵

• Anthropic’s high horse 🐴

• Botfestation 🤖

• AI 🤝 Environment

• China’s open source dominance 🇨🇳

• You are being persuaded by AI 🗣️

This newsletter is sponsored by… Me!

Create AI workflows and agents that help your team work faster and more efficiently.

Reply with “Agent” and let’s chat.

You’ll want to read this one in its entirety.

Enjoy 🍵

The West’s high ground

In a leaked memo, CEO and founder of Anthropic, Dario Amodei mentioned the dangerous necessity of taking money from Arab countries like Qatar, UAE and Saudi Arabia [Link].

Here is the full quote:

I find it funny when someone who works with the US government and has contracts with the DoD talks about the dangers of other authoritarian governments.

Anthropic has some of the best models on Earth. This is why OpenAI employees themselves use it, although Anthropic just cut them off from doing so [Link].

But this sentiment right here is why open source AI is so important. This belief that “only we are worthy enough to wield such power”.

Is Anthropic within their rights to cut off OpenAI from their products?

Of course. After all, Anthropic was literally created by former OpenAI employees who thought OAI was not taking AI seriously enough and would cause the end of the world.

But that’s not the point. The point is that at any moment, depending on how Anthropic feels, they can simply cut off access to intelligence at a moment’s notice.

In the coming world where we will have some sort of “super” intelligent AI system, who’s to determine what an AI can and can’t say?

When Anthropic and OpenAI say they want to align AI with human values, what values are these exactly? Who determines them?

If it weren’t for Zuck open sourcing Llama models and now Chinese AI labs open sourcing their models and research, we’d be living in a dystopia where a handful of companies could dictate what AI could say and couldn’t say.

The internet as you know it is already dead

New research suggests that 50% of the entire internet’s traffic is bots. I wouldn’t be surprised if it was even higher.

The reason why open source AI will be so necessary is because in the future, in fact in the present, every interaction you have with the internet will be through an AI system.

Do you remember when Google first launched the AI summaries at the top of a Google search?

I do. It was terrible. Everyone would make fun of them and rightly so. They were wrong and had terrible answers.

How about now?

Well, turns out they’ve gotten quite good. In fact, they’ve gotten so good that people don’t even bother reading anything else anymore. That’s right. New research from the Pew Research Center shows that people now don’t even bother clicking into web pages, and simply read the AI summaries and are satisfied with the answers [Link].

The study found users who see an AI summary are nearly half as likely to click on a traditional search result (8% vs. 15%). They're also much more likely to just close their browser after getting the AI's answer, ending their session right there (26% vs. 16% for non-AI results). And almost no one clicks on the links inside the AI summary itself; that happens in just 1% of cases. Essentially, the AI is satisfying the user's query directly, giving them little reason to visit the original source. A huge problem for anyone who relies on search traffic (hint: the entire internet), but that’s a whole other issue.

Is the AI that summarises the information biased? Who controls it? Has it been programmed with any ulterior motives?

Billions of people, literally the majority of the population of Earth, are using Google. Google controls the flow of information to most people on the planet. At the end of the day, Google is a company. Their allegiance lies with their shareholders. Not with truth or honesty. New figures suggest that ChatGPT now has over 800 million weekly active users. The level of influence these closed AI systems have over us cannot be overstated. And the dangers of allowing these closed systems to be the only access of artificial intelligence cannot be understated. With the level of influence these systems already have, frontier open source AI cannot come soon enough.

Mind you, none of this is even remotely exaggerated. Grok is an “unaligned” AI that can be accessed by millions of people online, and whenever it says something too unhinged, truth or not, it is censored and “fixed”. Considering truth is something that is subjective, it will be interesting to see how anyone can create a “maximally truth seeking” AI.

In other Anthropic news:

• A new report from Anthropic details what they think the US needs in energy to maintain a lead in AI. They estimate that the US must be prepared to run 50GW of power just for AI workloads by 2028. Anthropic proposes using the DOE's existing $5.75 billion in transmission partnership credit lines to fund AI-related grid projects, then selling those debt instruments to private buyers to free up capital for additional lending. Other proposals include launching loan guarantee programs for domestic manufacturers of critical grid components like transformers and circuit breakers, and expanding nuclear technology financing to meet AI's need for reliable base power. Anthropic predicts that in 2027, a single training run at the frontier will require 2GW and in 2028 it will be 5GW [Link]. For context, the US is looking into coal which was mentioned in an April 2025 executive order promoting coal as a solution for AI power needs, specifically directing federal agencies to identify regions where coal-powered infrastructure could support AI data centres [Link]. The funding for nuclear and the appetite for companies to invest in nuclear has also increased a lot with most major tech companies partnering in some kind of nuclear deal [Link]

• A federal judge in San Francisco recently certified a class action copyright lawsuit against Anthropic. Certifying a lawsuit means that a court has officially approved the case to proceed as a class action rather than just an individual lawsuit. The judge certified a class action representing up to 7 million copyright-protected books that Anthropic pirated. Mind you, these books are quite likely the high quality data Anthropic used to train their Claude models. This is actually a big deal because this could be very, very costly to Anthropic even if they settle. Highly recommend reading the article here [Link]

• Anthropic’s data company Surge AI somehow left out a list of websites they could and couldn’t scrape on behalf of Anthropic [Link]

AI is not killing the environment

Mistral has released the first-ever detailed report on the environmental impact of an AI model, and the results are fascinating [Link]. They did a full lifecycle analysis of their Mistral Large 2 model, and to the surprise of nobody who has been paying attention, the impact of AI is nowhere near as apocalyptic as some have made it out to be.

The main contributor to greenhouse gases and water consumption is in the initial training and ongoing inference, the raw energy needed to power the servers. But the actual usage for a single query is tiny.

Having their model generate an entire page of text, (about 400 tokens), generates as much greenhouse gas emissions as watching 10 seconds of an online stream.

The water used in that query is less than what it takes to grow a single small radish, and the raw material consumption is equivalent to producing a 2 euro cent coin. At least on a per-query basis, it’s really not that bad.

One interesting point is that Mistral believes location is key when building data centres. They’re building their data centres in France which means generally low-carbon electricity and a cooler climate, reducing the amount of water needed for cooling.

Guess where most large labs will be building data centres in the near future?

UAE.

We all owe China

I hope you’re not tired of hearing about China.

Since we last spoke, the Qwen team also released Qwen3-30B-A3B. A much smaller 30B model that is unbelievably good for its size.

The model only needs 33GB of RAM or CPU+GPU memory to run the model on 8-bit precision model at >6 tokens/s [Link]. Yeah 6 tokens a second is not a lot, but better than nothing.

This is the first time we have access to a model that you can run locally on a mac that is comparable to GPT-4o. This is the democratisation of intelligence.

In the last few weeks Chinese labs released Kimi K2 and then Qwen with all its variants. Now Zhipu AI (z.ai) have open sourced their latest GLM models and they’re looking pretty darn good.

As of right now, personally it’s hard to tell which is best between Kimi K2, GLM-4.5 and Qwen-3. I’d say Qwen is best for coding but for other tasks, I may give the edge to GLM-4.5 followed by Kimi. Really though, it’s all about testing models with different setups and prompts to see which is best for the use case.

The Qwen team also released Group Sequence Policy Optimisation (GSPO), the RL algorithm that powered the training for all their latest models which is really good for training MoE models [Link].

The only unfortunate issue with open source models is that their hosting options aren’t as good as closed source models. If I wanted to create an AI agent, it would be easier to do so with Claude and Anthropic’s APIs than it would be with an open source model. I’m currently building agents with different models so I’ll be writing a bit more about it soon.

Side note - z.ai is probably the best AI slide generator.

We’re not done either.

StepFun has also released their latest model, Step3 which is a 321B parameters MoE model [Link]. What I like about this model is that they’re trying new things technically; they’ve released a technical report with their model which you can read here [Link].

Alongside the model, they’ve also released:

• StepMesh - an open source communications library that makes serving large AI models split across multiple computers really fast [Link]

• StepFun-Prover Preview - AI that proves complex mathematical theorems step-by-step using formal verification code. The system writes formal mathematical code and verifies whether its proofs are correct in real-time. It uses over 1000 parallel verification systems to instantly check its work as it goes [Link]

Not to be outdone by all the LLMs, Wan AI released Wan 2.2, their latest AI video generation model. It’s the only open source model currently in the top 10.

For context, both Qwen and Wan are part of Alibaba group.

Not to feel left out, Tencent open sourced a first of its kind world model - Hunyuan3D. The model can create immersive and importantly, explorable 3D worlds from a single prompt or image. It can even export to Unity and Unreal Engine which is pretty cool.

I wrote about these types of models and what they’ll mean for entertainment and society in April 2023. I’ll be honest, I didn’t expect them to come this quickly. Other companies are also working on playable, explorable generated worlds. If you thought Minecraft and Roblox were big, wait till we have infinite world models that can be explored with near perfect graphics. Crazy times ahead.

To put into perspective what has happened in the last few weeks, China has open sourced:

• GLM-4.5

• GLM-4.5 Air

• Step3

• StepMesh

• StepFun-Prover Preview

• Wan 2.2

• Tencent Hunyuan3D

• Qwen’s GSPO

• Qwen3-30B-A3B

• Qwen3 Coder

• Qwen3-235B-A22B-Thinking-2507

• Qwen3-235B-A22B-2507

• Kimi K2 model + report

As of right now, Qwen models have surpassed 400 Million downloads globally and have spawned over 140,000 derivative models, surpassing Meta’s Llama models [Link].

History will look kindly on China’s stance on open source AI.

Meanwhile OpenAI still hasn’t released their open source model (rumoured to be releasing next week) citing safety concerns, Anthropic is significantly rate limiting their models and it’s possible we won’t get any open source models from either xAI or Meta.

In other open source news:

• Black Forest Labs has released a new state of the art open-weights text-to-image model designed to finally get rid of that generic, oversaturated "AI look" and produce more photorealistic images [Link]

• LG released EXAONE 4.0, a 32B model with a 131k token context window [Link]

AI Persuasion

A new paper called "The Levers of Political Persuasion with Conversational AI," has been released detailing how effective AI systems are at changing human beliefs. This is one of the largest studies of its kind, involving nearly 77,000 participants who conversed with one of 19 different LLMs, including frontier models (at the time) like GPT-4o. The AIs were tasked with persuading users on 707 different political topics, leading to a massive dataset where researchers then fact checked over 466,000 individual claims.

The study showed that how an AI is trained and prompted is far more important than the size of the model itself. This is rather important because this means someone can deploy a persuasive AI for cheap. For example, using a specific post-training technique on a small, open-source Llama-3.1-8B model made it as or even more persuasive than a much larger frontier model like GPT-4o. Highly effective AI persuasion tools can be built and used by literally anyone; if you’ve been thinking there’s more propaganda on the internet, you’re 100% right, there is.

What’s rather funny is that what we thought might be the most dangerous persuasion tactics, like personalisation and a common concern, were actually useless, never exceeding a 1 percentage point increase. Other popular psychological strategies, like moral reframing and deep canvassing actually performed even worse than a basic, non-specific prompt.

Unsurprisingly, the most effective strategy was simply assaulting someone with information. Simply overwhelming the user with claim after claim was 27% more persuasive than a basic prompt. I feel like this is a pretty good depiction of society today. People just screaming things at each other.

What makes this even funnier is that although this strategy was the most effective, it also led to the AI becoming significantly less factual. If the AI tried to be “maximally persuasive” irrespective of truth, it could shift opinions as much as 16 percentage points which is a lot. About a third of all claims the AI made when doing this were lies.

This persuasion best happens when there is a conversation taking place between the user and the AI. In fact, users were so impressionable that in a follow up study a month later, 42% of the initial attitude change was still present in the user. AI is creating a durable shift in belief. The study found this conversational approach was over 50% more persuasive than simply having a user read a static, AI-generated message.

Guess how most people use AI? …

Other news:

• Netflix is now using Runway's AI tools in its content production after it helped them speed up VFX work by 10x on a recent show. Disney is also apparently testing Runway's tech. This is a massive win for Runway and honestly a fumble by OpenAI, which was thought to be a frontrunner to land a deal with Disney for its Sora model. Sora is nowhere near the best video model right now, although it’s rumoured the v2 version is coming very soon [Link]. Btw, they used AI in the Netflix series “The Eternaut”.

• A new 3B parameter model called ColQwen-Omni can embed 30 minutes of audio in just 10 seconds and perform multimodal search across documents, audio, and video without needing to transcribe the audio first. Super efficient retrieval for mixed media Link

• A new open-source model called MiroMind-M1 uses new training methods to increase the performance of models in mathematical reasoning. It’s already posting SOTA results on some math benchmarks. The entire project, including datasets and training configs, has been released to the public [Link]

• ByteDance Seed Prover got a silver medal at the IMO, correctly answering 4/6 questions [Link]

• Harmonic, a company looking to build mathematical superintelligence, also released their results on the IMO, getting a gold medal and correctly answering 5/6 questions in the IMO [Link]. They’re also releasing Aristotle on the app store, which can help with math questions and proofs.

I want to keep bringing these AI updates to you, so if you’re enjoying them, it’d mean a lot to me if you became a premium subscriber.

As always, Thanks for Reading ❤️

^{Written by a human named Nofil}