Just say no to OpenAI and Anthropic

I just lately learn End Times Fascism and the Fight for the Living World by Naomi Klein and Astra Taylor. It’s a wonderful ebook and I extremely suggest that everybody learn it. Tech oligarchs make up an essential element of the end-times-fascist coalition. One factor that stood out to me, as a software program and safety engineer who finds LLMs to be extraordinarily highly effective, is that the American frontier labs are run by unhinged lunatics who genuinely consider they’re creating the Machine God, that they are effective with burning down the globe to result in AGI, and that they consider with spiritual fervor that they are the one ones good sufficient to securely management it. I already type of knew this, however this ebook actually made it sink in.

As a lot because the AI hype throughout us pretends that OpenAI and Anthropic are distinctive and particular, they don’t seem to be. They’re not the one firms doing this work. In reality, a whole lot of the open weight fashions are nearly pretty much as good as the most recent ChatGPT and Claude fashions, together with for extremely advanced duties like vulnerability analysis.

Unlike fashions from OpenAI and Anthropic, open weight fashions have main benefits:

• It’s attainable to run them privately by yourself {hardware} (when you can one way or the other get a pc with sufficient RAM)
• You can use a supplier like Tinfoil that depends on safe enclaves to run them privately (I speak about this in a bit extra element under)
• If privateness is not an enormous concern, you need to use any variety of different suppliers, a few of them fairly inexpensive, and it is easy to change so you are not locked right into a single evil vendor
• You can nonetheless use highly effective LLMs with out financially supporting deranged oligarchs who gamble with life on Earth in order that they will at all times stay in energy

ChatGPT and Claude will by no means be non-public

These firms solely provide closed variations of their highly effective fashions (regardless of the phrase “open” in “OpenAI”). You pay them for chat and/or API entry, and they’re allowed to have a look at every little thing you do.

They provide a checkbox so you’ll be able to choose out of them utilizing your knowledge for coaching, however your chat periods nonetheless aren’t non-public. They have entry to all of it. They can have a look at it, share it with governments or firms, or do no matter they need. It’s attainable, probably even, that they will not use it for coaching when you’ve opted out, at the very least not proper now. But do you actually belief that long-term? Have they executed something to make anybody consider that they respect individuals’s knowledge?

OpenAI and Anthropic are company gatekeepers

Their fashions have strict guardrails, as anybody who’s tried to make use of them for something cybersecurity-related has little question observed.

Every session has a cop agent studying your transcript over your shoulder, ensuring you are not utilizing it for hacking. If it thinks you is perhaps, the cop agent stops your session from persevering with. It’s very strict, and false positives occur on a regular basis. Of course, not everybody has the cop watching them.

Some individuals get entry to the fashions with looser restrictions or with none in any respect. OpenAI has Trusted Access for Cyber and Anthropic has the Cyber Verification Program. To acquire entry to extra succesful, much less policed variations of their fashions, people and firms want to use and be accredited by OpenAI and Anthropic. A variety of US cybersecurity firms get entry. Almost definitely US navy contractors do too, and doubtless any elements of the US authorities that need it.

While there’s some benefit in proscribing highly effective fashions, handing the last word energy to determine who will get entry to them to Sam Altman and Dario Amodei – with session from the Trump Administration, after all – is a horrible, undemocratic thought.

I believe a greater strategy is to easily not put your self in a state of affairs the place you depend on ChatGPT or Claude. Instead of begging company overlords for privileged entry, it is higher to spend your time determining find out how to get fashions like GLM-5.3 to fulfill your wants. And it is not onerous. These fashions are good, with the additional benefit that they do not have a cop watching over your shoulder!

Open weight fashions are spectacular

So far, open weight fashions appear to be about 4 months behind the flamboyant proprietary fashions from OpenAI and Anthropic.

Four months is nothing.

I’m extraordinarily impressed proper now with GLM-5.3. There’s loads of knowledge and benchmarks on the market, however my vibes-based opinion from utilizing it’s that it is mainly on par with Claude Opus 5.5 for software program engineering.

I’m additionally impressed with GLM-5.3-Flash. This is just like Claude Opus 4.8. It’s a lot smaller and cheaper than GLM-5.3, and nonetheless extremely succesful. It’s additionally truly possible for a person particular person to get {hardware} to run it, like when you can someday get your arms on a Mac Studio with 512GB of RAM.

GLM fashions are made by a Chinese firm referred to as Z.ai. While the fashions are spectacular, I do not suggest making an account instantly with Z.ai or sending them any of your knowledge. Their coding harness ZCode (their model of Claude Code) was recently caught silently uploading native supply code repos and sending them again to Alibaba Cloud with out consent. Also, it is secure to imagine that Chinese firms will share any knowledge you present to them with the Chinese authorities.

Other open weight fashions from Chinese firms which can be spectacular embody the DeepSeek, Qwen, and Kimi households of fashions. It’s completely effective, secure, and safe to make use of these open weight fashions. Just do not belief the businesses themselves along with your chat periods.

Another frontier AI lab price a glance is Mistral. As a French firm, it is the one huge participant that is not Chinese or American. They simply launched an open weight frontier mannequin referred to as Mistral Large 4, aka Le Chonk. I believe it is fairly a bit extra affordable to belief Mistral along with your knowledge in comparison with the Chinese or American firms.

Private inference with Tinfoil or Maple AI

It’s attainable to have non-public AI, even with out internet hosting it regionally your self, utilizing safe enclaves. These aren’t excellent, however I do assume they supply a really significant distinction in privateness in comparison with the established order mode, which is “the supplier can entry every little thing with no restrictions.”

The firm that is doing essentially the most work right here is Tinfoil. If you need to study the main points, learn their docs on safe enclaves and distant attestation. Another firm doing that is Maple AI – though, technically, Maple makes use of Tinfoil for inference within the background.

Both Tinfoil and Maple are good choices. Tinfoil is cheaper, however Maple is extra consumer-focused.

If you need a ChatGPT-like expertise, the place you’ll be able to discuss to a chatbot, have it look stuff up for you on the web, give it entry to docs to learn or do stuff with, entry it in a cell app, and so forth., then I’d go together with Maple. There’s a small free plan and a Pro plan that prices $20/month.

If you need to do actual agentic coding duties like software program engineering, reverse engineering, malware evaluation, CTFs or pen assessments, and so forth., then you need to use Tinfoil as a substitute. They cost by the token for personal inference, so no month-to-month plan. GLM-5.3 at the moment prices $1.80 per 1M enter tokens, $0.80 per 1M cached tokens, and $5.75 per 1M output tokens. You can use considerably fewer tokens when you select the “low” reasoning stage, and it is nonetheless very spectacular.

Inference with out the privateness

If you do not need to pay a premium for personal inference from Tinfoil or Maple, however you continue to need to free your self from OpenAI and Anthropic, there are tons of less-private choices which can be additionally very low cost.

One that I’ve been enjoying with is OpenCode Go. It prices $10/month for greater than $10 price of credit. This is an inference router, which mainly means your API queries do one thing like this: your pc -> OpenCode Go API -> the precise inference supplier. This signifies that each OpenCode Go and no matter backend supplier they’re utilizing may have entry to your session knowledge. That stated, the privacy section of their docs guarantees which fashions knowledge will not be used for coaching (most of them) and what the information retention is (zero days for many, together with all of the GLM, Qwen, Kimi, and DeepSeek fashions). Of course, it is only a promise. Nothing technical ensures it.

There are loads of different low cost choices on the market for non-private inference, when you simply need to experiment with highly effective open weight fashions.

How I’m utilizing these fashions

I’ve been utilizing each Tinfoil and OpenCode Go, configured with my present harness of selection, OpenCode. I’ve been principally utilizing GLM-5.3 on low reasoning, with nice outcomes. (Things change so shortly with the open weight fashions, who is aware of what will likely be greatest six months from now.)

I exploit OpenCode Go inference for utterly public work, like when I’m implementing one thing that is already open supply.

I exploit Tinfoil for every little thing else.

I’ve solely simply began working this manner, however to this point, I believe the open weight fashions can simply utterly substitute my use of ChatGPT and Claude, at a comparable or cheaper value, with the additional benefit that they are much more non-public and there are far fewer guardrails.