...
I think Open source LLM's will hit a ceiling for this one reason: safety guardrails. Today we see Mythos and GPT 5.6 Sol put under heavy scrutiny for the primary reason that it is too unsafe to release to the general public. The guardrails come in three layers - safety baked into the model itself, immediate flagging and offline batch analysis. Level 1: Baked into the model Here's a strange example…
I often hear that AI is resulting in 10k LOC reviews and this is creating a bottleneck. I don't think you should waste time reviewing every single line of code in here and just use AI to review it! What you contribute as a reviewer You need to know what you contribute as a reviewer. As a reviewer, you contribute your Out Of Distribution knowledge that the author or the LLM might not have Its a…
I like the UX of speaking to voice assistants and getting stuff done. Usually it involves something not so trivial and I do expect intelligence from the assistant. But the current state of AI voice assistants is bad! Either they are stupidly slow and unusable or quick but stupid. BUT. There's a clear winner. And the winner is Grok - I'm not going to hide it from the rest of the blog post. I'm…
MCP is cool but it needs a verified human in the loop approval button. Here's what I mean in a picture. The reason this is important is because there must be a cryptographically verified way for the server to guarantee that it showed you the payload and that you the human have approved it. If the MCP is set up in such a way that the specific method is gated behind human, there is no way for the…
No one knows this but different LLM apps are allowed to access different kind of real time data sources. The obvious ones are obvious: Grok allows you to search through Tweets and ground answers based on them. This is a real moat because X doesn't allow other companies to query Tweets real time. I find this interesting because what data sources are allowed from which LLM application also adds a…
I like the UX of speaking to voice assistants and getting stuff done. Usually it involves something not so trivial and I do expect intelligence from the assistant. But the current state of AI voice assistants is bad! Either they are stupidly slow and unusable or quick but stupid. BUT. There's a clear winner. And the winner is Grok - I'm not going to hide it from the rest of the blog post. I'm…
In Europe each country has its own idiosyncratic public transport presence online. Netherlands has NS and NS international. Both are different apps, one works internally and one for international. There's also GVB, 9292 and OVPay but I forget what's the best way to do it (this is kind of my point). All of them have some overlap. To be fair - Netherlands is one of the simpler countries that just…
The last two decades of (desktop) software engineering was mostly defined by bloated software like Electron, resource hogs like Chrome and outright slow and buggy UI. Not to mention slow programming languages like Python, Ruby and Perl. To be clear: these developments were not unexpected or bad - they allowed us to create way more products and iterate quickly so it was a net positive. Electron…
I often hear flat out false statements about the state of LLM economics, one of them being that the costs are "unsustainable". This is blatantly false and easily falsifiable with just a few minutes of thinking. Of course I can never know what really goes on internally in companies like OpenAI and Anthropic but I feel I can make really good guesses. I'll lay out a few proxies that can help drive my…
I'm a big fan of SKILLS.md in the context of LLM agents. On the surface, its a very simple concept - it just contains summary of a big concept and this summary is always fed into the LLM's context. It can lazily load the full concept when required - it can know exactly when to load it since it always has the summary inside it. What Agent skills miss I think Agent skills are good but they need two…