AI

Alternatives to: ChatGPT, Claude, Gemini, MetaAI, Grok

Much like with browsers and search engines, everything you do in Big Tech AI products is recorded, stored in a creepy profile about you, and read by humans. Never put any personal info into a Big Tech AI chatbot!

The problems of AI go much deeper than that, however. The development of the 'frontier' models has concentrated power in the hands of a few companies. And those companies are all backed and part-owned by Big Tech companies. So by using the mainstream AI products you are entrenching Big Tech's monopoly power. Furthermore, the environmental impact of AI data centers is considerable. Their models were all trained by stealing copyrighted works. The models contain bias, and you'll never really know how bad that is because they are closed source. Their chatbots are weighted towards sycophancy and persuasion. Prolonged use can lead to cognitive decline and in some cases psychosis [1] [2] [3].

That said, the underlying technology has some clear use cases. These include coding, cybersecurity, and (if you force the model to include citation web links) some research tasks.

The ethical problems with the training methods, and the environmental effects of data centers, cannot be fully mitigated. Therefore your decision making process should be:

Decision process for AI

1. "Do I really need AI for this task?". In many cases the answer is no. However, if yes, move ot step 2.

These days many people use AI for basic web search. This is computationally much more expensive and therefore uses anywhere between 10 times and 60 times more energy than regular search engine search. So just search normally!

2. "Could I get what I need from running an open source model on my own computer (self-hosting)?" This is easy to do so is largely determined by whether you have a powerful enough computer. If so then skip down to the second table on this page, Self-hosted AI. If not then:

3. "Which hosted provider is non-Big Tech and protects my privacy?" Choose providers from our table below, preferably with "Yes" in the Zero Knowledge* column.

4. "Which providers have other ethical features such as ethical training sources, and using their own infrastructure?" Choose from the info listed in the 'Ethical Features' column of our table.

5. "Do I really need a massive AI model for this task?" Bigger models are more expensive to run both in terms of costing you more money and costing the environment. For many tasks you can use the 'Quick' option (instead of 'Thinking') and you can choose smaller models. For example the more ethically trained models are less powerful, but sufficient for many tasks. Imagine you're going down to the shops and you have a choice of cars - you're not going to choose a monster truck when a Ford Fiesta will do the job.

Hosted (Cloud) AI providers

*Concept definition: "Zero knowledge". This means a combination of End to End Encryption (E2EE), and secure hardware in the pathway that the service provider cannot access (called TEEs, or Trusted Execution Environments). The TEE is where the model processes your prompts.

Provider name	Models Offered	Zero Knowledge	Ethical Features	Business Model	Country of Origin
Maple	Open source only, incl. GLM, Kimi, Deepseek, Gemma (you choose)	Yes	E2EE - your data is entirely private and not used to train models either; The most powerful open source models; They host the models on their own racks; Proxy API allows you to use it in OpenCode	Free tier (two basic models, less chats, no web search); Paid tiers $20 and $100 (better models, higher rate limits)	USA
Tinfoil	Open source only, incl. GLM, Kimi, Deepseek, Gemma (load-balance or you choose)	Yes	E2EE - your data is entirely private and not used to train models either; The most powerful open source models; They host the models on their own racks; You can host them in containers	Three free questions a day; Paid tier $20 (Up to 20M input / 1M output tokens/hour) or API (pay as you go)	USA
Confer	Open source only, but they don't disclose which. They load-balance between models (you don't choose)	Yes	E2EE - your data is entirely private and not used to train models either; The most powerful open source models; They host the models on their own racks; FYI login requires a passkey; Founded by the founder of Signal	Free tier (one basic model, 25 questions per day); Paid tier $35 (better models, higher rate limits)	USA
Duck.ai	Big tech and Open source models	No, but uses the proxy model to hide your identity from the model provider	Your sessions are anonymised; Models are not trained on your data	Free, with contextual ads at the top of search results	USA
Lumo	Chooses from open source models in the background e.g. Qwen, GLM and Apertus (you don't choose)	Yes	E2EE - your data is entirely private and not used to train models either; They host the models on their own racks.	Free limited tier; Paid tier $10 (more use of best models, unlimited chats)	Switzerland
Euria AI	Uses the open source models Qwen and Apertus	No	Owned by Infomaniak (and can be part of their kSuite); They promise your data is entirely private and not used to train models either; They use green energy	Free chats on their site; Included as part of kSuite	Switzerland
Thaura AI	Uses the open source model Qwen	No	Syrian founders; They promise your data is entirely private and not used to train models either; Data centers in Finland; Claims to be relatively unfiltered	Free tier with limited chats; Paid tier $13 unlimited	Finland
Green PT	Open source only, incl. GLM, Kimi, Deepseek, Qwen	No	Attempts to be 'green AI'; They promise your data is entirely private and not used to train models either; Data centers in EU;	Paid tiers with free trial	Netherlands
Empathy AI	Uses the open source models Qwen and Olmo	No	They promise your data is entirely private and not used to train models either; They host the models in the Net-Zero data center in Asturias, Spain; very anti-Big Tech manifesto; Designed to be used on your own data (query your knowledge base)	Free chats on their site; Paid offering for querying your own data	Spain

RTA's choice

Our team uses Maple, Duck.ai and Tinfoil. And self-hosting. All for software and web development.

Self-hosted AI

This means you are running AI models stored on your own computer (locally). No need for the internet and no data gets sent out anywhere. To run large models you need quite a powerful computer, but for smaller or more optimised ones you can get plenty of tasks done with local models.

Definitions: App = software to access models in a chat interface; Harness = software to use models to perform actions on other files e.g. coding; Model = a large language model you can download

Provider name	App / Harness / Model	Features	Business Model	Country of Origin
LM Studio	App	Easy to install and use; Search and download any open source model; They also provide an agent to use with your local models; Hosted options are zero knowledge	Free to self-host models; Paid hosted model plans	USA
Ollama	App	Less intuitive UI than LM Studio; Use any open source model but cannot search them from the app	Free to self-host models; Paid hosted model plans	USA
GPT4All	App	Simple UI; Search and download any open source model; Works well on less powerful hardware	Free to self-host models; Paid plans for access to agents	USA
OpenCode	Harness	FOSS; Not intuitive to set up; Allows you to have the AI act on the underlying files you point it at; Some free models available; Can use your local models, or a paid API e.g. Maple	FOSS; Paid tiers to access hosted models	USA
Apertus	Model	Fully open - open data, open weights, open research; Provided by the Swiss AI Initiative	Free; Access via HuggingFace or LM Studio	Switzerland
Olmo	Model	Fully open - open data, open weights, open research; Provided by AllenAI, a non-profit	Free; Access via HuggingFace or apps like LM Studio	USA

Next up:

➡️ Laptop OS alternatives