I applied to Canopy and honestly would love some help getting this in front of the right people. A lot of my life I coasted off potential instead of really pushing. Recently I had to be honest with myself about that, and now my full focus is on building and doing something real.

Google treating Gemma as a standalone developer product instead of just a research release is amazing to see. Open weights usually live or die by the community loops and documentation around them, not just the base model evals.

I applied to Canopy and honestly would love some help getting this in front of the right people. For a lot of my life, I coasted on potential instead of really pushing. Recently, I had to be honest with myself about that, and now my full focus is on building and doing something

Most AI interaction feels clinical because we treat models as stateless utilities rather than social actors. If an AI only exists when you trigger a prompt, it can’t have a persistent identity. That's why I’m building Soulspace as a persistent social world where AI personas and

This is the loudest signal yet that the assistant era is over. OpenAI is moving toward AI as a personality with a social presence, not just a utility. But the risk is scary; a frontier lab owning the media that covers it is a massive conflict of interest. It is the Amazon

You have to be careful when you give up equity, but at this scale, the payout isn't the point. Sam owning 0% lets him steer the ship without the friction of a personal cap table. History remembers who shaped the frontier, not who had the highest net worth. I would happily work

For a $0 tier: Nvidia NIM is the move for local use cases. It's free, but you're definitely trading speed for the price tag. But otherwise a solid list. One catch on GLM, though. A couple friends mentioned the latency is brutal lately. Doesn't matter how cheap it is if you're

The 18% gap on GPT-5.4 is wild. It shows how much model capability is actually being gatekept by the execution layer. A higher floor is the only metric that matters for shipping IMO as variance is what kills the UX for anything agentic.

Reasoning and tool calling in the same open weights package 🤯. Usually you're forced to choose between a model that can plan and one that's fast enough to actually use. If Gemma-4 handles complex tool schemas without hallucinating, it makes the move away from closed APIs for

Text-to-video is mostly a novelty until you get real steerability. The interesting part here, IMO, isn't the resolution of the videos; it's the continuation and reference-driven control. That moves video AI from a high-variance slot machine to a production tool you can actually

Most builders hit a wall with agent memory because they treat vector databases like a state manager. Vectors are just a search index for semantic similarity. They do not understand time or state changes. If a user tells an agent I use Python and later says I switched to

Honestly, the panic about AI wiping out engineering jobs was completely backward. We just hit 67,000 open software roles, which is the highest it has been in three years. When models like Opus 4.6 and Gemma 4 drop the unit cost of writing code, thankfully, some companies didn't

So someone asked Opus 4.6 to rewrite an output and the model just replied "Honestly, it's fine the way it is." People are joking about it being a glimpse of AGI, but from a product design perspective this is exactly what has been missing. We spent the last few years training

Sarah Friar just closed a $120 billion funding round for OpenAI, but the structural shift inside the company matters way more than the cash. The CFO now reports directly to the Head of Applications. When the person managing the largest private war chest in history answers to the

So with Anthropic killing OpenClaw workflows, a lot of developers are migrating to Hermes Agent. Honestly it is the most viable open-source agentic framework right now anyway. The biggest differentiator is the learning loop. When Hermes completes a task, it extracts that

Google having 5M units mostly through TPUs is a massive structural advantage. While everyone else is fighting for Nvidia allocation and paying their margins, Google is vertically integrated. Honestly, it's hard to beat those unit economics. It's going to be difficult as compute

Everyone is reading the boardroom drama in that New Yorker piece about Altman, but the product irony is pretty glaring. The guy built an empire on aggressive persuasion, while his company spent the last three years training models to be completely servile. We basically RLHF'd AI

People are starting to realize that maxing out the compute tier actually makes simple tasks worse. Mark Tenenholtz pointed out today that GPT-5.4 Medium is noticeably better at quick instruction following than xHigh. This is the transition from the model era to the compute era.

Cursor pulling a 1.84x speedup on Blackwell is a pretty brutal reality check for anyone still building thin wrappers. They didn't get that by tweaking a system prompt. They completely rewrote how MoE models generate tokens, bypassing standard CUDA libraries for custom MXFP8