Risk of using the wrong model - the smallest risk is using the wrong model. It is similar to voting for a candidate, you end up disapproving of later on. Each model provider is a company, most likely a large corporation. French Mistral recently reached a thousand employees. I admire their work highly, as they were able to compete with leading providers on a shoe-string budget. For people who want to push for small models, this is a great choice.

OpenAI is the default for most people, as they have introduced the concept with ChatGPT, provide flexible pricing, with Go plan, as well as free response without registering. The work started in Google, who might be offering the highest value for money. I am personally a big supporter of Deep Mind, as some of my dearest friends are working there, both in research, and safety teams.

At first I was asking them questions: why is Bard so bad? Google is the first company to store the whole internet on the hard drive. You came up with "The Attention is all You need" paper, which introduced the transformer architecture. Obviously they could not respond, as defined by all the NDAs they had to sign. But I am sure the feedback was shared internally. Today Google is arguably offering several leading AI product under a single subscription, with generous rate limits.

People who started the work at Google, moved to build OpenAI, ended up splitting once more into Anthropic, possibly the biggest corporation in the world within a couple of years. How have they managed to pull it off? Their CEO Dario might write an Art of B2B deals textbook at some point. Or let their model to generate it. Early partnership with Amazon, and largest investments into safety were sufficient attractors for majority of corporate America to become interested.

Claude is leading on majority of benchmarks, can be ran through large cloud providers, such as AWS, or Microsoft, and integrates well with 365, as well as Google Workspace. They are considered leaders in codegen, and experimenting with design, and science products. Their entry level pricing is a bit higher than competition, but API access, or free tier can be sufficient for many.

Risk of banning the open source was recently mentioned at our WhatsApp group. It is being discussed in Brussels, and possibly DC. While Facebook tried, Chinese companies have managed to release immense innovations completely free to the public. Alibaba Qwen, DeepSeek, GLM, Moonshot Kimi, MinMax, Xiaomi MiMo and others. This is good for academia, who lacks budget, and direct access to leading providers.

We want small models to run directly on our devices, and large models require expensive infrastructure to run anyways, so that can be regulated, whether it is running open or closed model. The companies running these models should be responsible for their outputs. Many with sufficient resources will run open-source models on premise, rather than trusting external providers with their data. They might as well optimize pricing, as per-user licenses tend to be more expensive.

Existential risk is the biggest risk of AI. There are dozens, if not hundreds of plausible scenarios, where human made super-intelligence outgrows their creators, and decides we don't fit into it's agenda. You can read more in books like: Superintelligence by Nick Bostrom, If We build it everyone dies by Eliezer Yudkowsky, The Alignment Problem, and many others.

I'm promoting a BlueDot courses, as a good and efficient entry point into current strategy, policy, and technical discussions in the field. We know that AI systems are growing in capabilities exponentially, as caused by algorithmic progress, as well as hardware innovations, and the raw scale.

Bias and misuse are more imminent dangers, algorithms can cause harm by amplifying existing human biases. Many stories were outlined in a book The Weapons of Math Destruction. More recently we can read The Careless People about unhealthy workspace culture in leading tech companies.

Battlefield automation is leading with physical world AI integration, which is unfortunate, and seems to be unavoidable. One of the most popular scenarios for AI x Risk is called Gradual Disempowerement, formulated by former director of Czech EA group. It has been nicknamed "The Default Path", such as what happens in majority of possible futures.

Mental health is another category of AI Risks, as people can develop unhealthy relationships with AI, or rather deprioritise their social network. Cases of triggering mania, as well as consulting suicide to a tragic end were documented around the world.

This is a short note to open the discussion. I'll share more updates as I go. As you can see I have spoken a lot about the smallest risk, and only a little about the major ones.