Accelerando is a true masterpiece. Crypto and endless speculation, AI and lobsters, space exploration - all in all just "this is our near future". TBH, I know of just handful of Sci-Fi novels as fundamental and as let's say prophetic as this one. The others to my taste in the same category is Nexus trilogy by Ramez Naam (even if a bit farfetched by now), The Diamond Age (="The Illustrated Primer" is peak AI) by Neil Stephenson and Daemon+Freedom by Daniel Suarez (=AI + crypto DAOs).
I’ve been thinking a lot about how AI assistants evolve—whether they’ll keep absorbing specialized tools (like ChatGPT making search engines redundant) or if there’s another path. Model Context Protocol (MCP) feels like a turning point.
Most people talk about MCP as just a tool-calling API, but I think that misses the bigger picture. It could redefine how AI assistants work—not as standalone products but as dynamic platforms orchestrating external AI services.
In this article, I explore MCP’s implications through four angles:
- AI Competitive Dynamics – Why generalist AI models tend to absorb specialized tools.
- Economics of Complements – How AI assistants can thrive by enabling third-party tools instead of replacing them.
- MCP as a Personal Agentic Platform – The shift from chatbots to deeply integrated AI assistants.
- MCP as an Agent-to-Agent Protocol – Why AI systems should talk to each other instead of relying on humans as intermediaries.
I’d love to hear from others experimenting with MCP — does it feel like just another API, or is there something bigger here?
So far, the most interesting part is "Link diagrams to let users drill down to details or browse related subjects". How do you do that? How does it actually look? The sign in process is prohibitively complicated, sorry.
Apart from this feature, looks like a glorified and confusing draw.io.
User side of false negatives: You miss skin cancer. The user delays a visit to a doctor by a month. The user dies in 3 months, and her relatives sue the hell out of the model creator.
User side of false positives: The model thinks a blemish is malignant. The user spends few grand to verify it is not, and blames you for scaring her.
Doctor side of false responses: The fucking engineers do not know what they are doing. Please, my patients, do not use that. We the doctors and the responsible patients should unite against the stupid AI.
Arguably, the user's side is about being moral. The doctor's side is more important for adoption.
BTW, what is the accuracy, false positives and false negatives for just an AI model, just a doctor, and a doctor equipped with an AI helper model?
Ahhaha, Intel's TDP is 2-4 times bigger under full load (and I guess ~10-15x under moderate load) while having negligible perf advantage over M2. The author can not be serious.