https://simianwords.bearblog.dev
12 posts
Tech
Subscribe via RSS
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 modelHere's a strange example from Sonnet 4.5 Level 2: Immediate flaggingMythos was so cautious that you were not able to ask it…
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 reviewerYou 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 mistake thinking you can outsmart an LLM into nitpicking few lines of code here and there. This is not…
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 assessing across these dimensions intelligence speed voice reliability and accuracy can I interrupt and…
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 so that the server knows it showed you the payload and you have approved it. So it needs two things the payload your cryptographically verified approval If the MCP is set up in such a way that the specific method is gated behind human, there is no way for the agent to make changes on your behalf no matter…
No one knows this but different LLM apps have different allowed data sources to ground the answer with. 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 lot to the utility. RedditReddit at least used to be a good source of believable anecdotes and a fair…
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 requires you to tap in and tap out using Debit card and this process generally works well. The other…
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 abstracted out nitty gritty's of handling UI in specific OS contexts and no one wants to deal with…
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 point. Note that I'm not really talking about profitability of these companies. I'm only speaking…
