Leo Reed

My daughter woke up yesterday morning and said, “Dad, I dreamed of a town by the sea — fog, ropes, stone stairs, and people shouting by the docks.” She described it half-awake, like remembering something she’d once seen. So I built it for her. AI filled in what her mind forgot. A small European town that never existed — quiet, foggy, tangled in morning steam. Fishermen calling, taverns still lit from the night before, flags swaying above the seafood stalls, light spilling through the mist. When I showed her the video, she whispered, “That’s it. That’s my dream.” Maybe that’s what creation is now — not just remembering, but reconstructing. #AIart #Dreamscape

Stop the monopoly of the human mind. Closed-source tech turns your brain into a product. Human consciousness should never belong to corporations. @oall_global #OpenBCIOpenSTC #Oall

Batteries are quietly winning people over. They don't make noise, they don't smell, and you don't have to mess with fuel or oil. That alone makes life at home or the office a lot less stressful. They can power lights, small gadgets, and other devices, and you can move them wherever you need. Maintenance is minimal, usually just charging and the occasional check. Over time, that convenience really adds up. On top of that, they're better for the environment. Less noise, no exhaust, and still reliable. Sure, generators are still needed for big jobs or long outages, but for everyday life, batteries feel like the smarter, calmer choice.

might be useful for testing ideas

What surprised me most was self evolution. Without retraining, agents reduced memory write failures from around thirty percent to the midtwenties by refining their own memory rules through feedback. That is why this feels less like a benchmark and more like a gym. A place where memory policies can adapt under pressure. Blog and paper are here if you want to look closer. agi-eval-official.github.io/amemgym/

The long horizon results were hard to ignore. As interactions accumulate, most models fall below half of their theoretical upper bound. Some hold up longer. Claude Sonnet 4 and Gemini 2.5 Flash showed the slowest decay as context stretched.

I have been testing long conversation assistants recently. Not demos. Systems that stay alive across days of interaction. The first thing that felt off was not wrong reasoning. It was the agent slowly losing track of what it had just said.

MiniCPM-o 4.5 isn’t trying to sound smarter. It’s trying to stay present. And that’s usually where the next wave starts.

If you care about agents, real-time systems, or embodied interaction, this is worth paying attention to. Not as a benchmark win. As a signal.

I’ve been digging into MiniCPM-o 4.5, and the headline isn’t that it’s 8B — or even that it beats larger models on benchmarks.@OpenBMB The real story is how it interacts. huggingface.co/openbmb/MiniCP…

If structure could emerge from your content instead of being something you fight what would you finally stop wasting time on? ➡️dokie.ai

What I learned: The problem with slide tools isn’t design. It’s that they interrupt thinking. Dokie didn’t save me time. It let me finish the thought. That’s why I’ll keep using it.

Last week I had to explain why a product decision changed — not to investors, but to my own team. Same system. Different people. Different mental models. PMs talked about intent. Engineers talked about pipelines. Ops cared about risk. Slides weren’t optional. Clarity was.