The arrival of OpenClaw at the end of January and the reveal of NemoClaw on Monday at GTC have made for a busy few weeks for personal AI. OpenClaw is now the largest and fastest-growing open-source project ever. Jensen Huang of Nvidia declared, “OpenClaw is the most important software release probably ever,” on March 4 and recently told Jim Cramer, “OpenClaw is the next ChatGPT.” Another really surprising reveal was Nvidia/Nemotron-3-super-120b-a12b, and the support for open-source models by companies not from China. The question of how any of this can be useful is on many people’s minds as they watch this news. Why all the excitement?
Deepseek-R1, released just a year ago, has been overshadowed by ChatGPT, which has been available for over three years. Microsoft has been offering Copilot to its users for some time now, and this year it added it to its E5 licensing for enterprise users to use in its Studio product. Similar to how Chrome integrates Gemini, Firefox incorporates Perplexity, and Codex and Claude Code gain browser canvas control, Twitter/X has Grok, and Facebook has some kind of Meta model in the background, Microsoft has also now integrated it into Edge. While some argue that AI is being “forced down our throats,” others find it useful and anticipate discovering even better applications in the future. I haven’t mentioned Apple with Safari yet, but I’ll get to that.
Developing a balanced AI demands substantial effort and billions of dollars. Ensuring an AI model’s bias remains as neutral as possible is essential. While I appreciate the transparency of Chinese models, which publish their weights for public analysis, anyone familiar with the multitude of Chinese models knows the CCP bias is obvious. Open-source AI models had a promising start in 2024 with a “Cambrian Explosion,” but they seem to have disappeared around seven months ago as the competition to achieve AGI intensified. No one wanted to keep the recipe for their secret sauce out there for all to see if someone else could just take it and mix up their own brand of seven herbs and spices. The alleged “mass data theft” of Anthropic and OpenAI’s knowledge through distillation by Chinese companies has further increased the secrecy. The arms race is on! Europe is realizing that relying on American AI may not be the best option and is urging Mistral and its allies to secure additional funding. Russia closing off its internet and creating its own Yandex and ruLLaMa models is just a clear attempt to fence off its own sphere of AI influence. Having biased models like xAI and Grok isn’t ideal for anyone, and in the US, it only deepens the political divides we already face. Having sovereign open-source models is existentially vital, so other countries will develop their own AI bias, which gets labeled “neutral” to preserve their values. Things are getting more conservative and preservative in the face of all this competition. Apple, the US’s homegrown computer giant, is nowhere to be seen in this.
So, Nvidia announcing last week that they will spend billions on open-source models is a significant development. Sure, Nvidia promised billions to fund other AI companies under Trump’s AI initiatives, including Stargate for data center buildouts, but much of that has not materialized over the past months, as companies like Anthropic and OpenAI have fluctuated in value based on the rants of a former Fox & Friends host, and questions about the feasibility of powering so many of these data centers in the short term come up. The GTC Preshow conversation with OpenClaw creator Peter Steinberger and the subsequent reveal of NemoClaw during Huang’s keynote have brought personal AI to the attention of many more people outside the AI evangelist community. You can’t have gone through the past 20 years of computing without hearing that name, as their GeForce products were integrated into gaming and home computing systems. (AMD Instella who?) Now, Nvidia is stepping into the world of tomorrow, with something you can experiment with on any old computer. What’s the big deal if you can just ask ChatGPT or Claude? The difference is personal AI sovereignty.
None of this is free. If you’ve never used Claude or Codex for work, beware that they can be addictive and costly. I’m not talking about the Copilot or ChatGPT models you get for free. The fact that those agents have access to the entire Internet hobo soup of information and misinformation means hallucination is built right into the feature set. I’m talking about the real AI that can do deep thinking: the Pro versions cost ~$20/month and offer more thinking power and more tokens, and the Premium versions for builders and creators range from $100 to $120/month. Those interested in local models aim to run on hardware such as a DGX Spark, a Mac Mini M4, or even a Mac Studio to handle queries without relying on cloud providers, which pose privacy and safety concerns. The models capable of running on such devices are very powerful, especially if you connect four maxed-out Mac Studios with shared memory. Does this mean cloud services are in trouble? Not at all. Like I said, none of this is free. How is Apple going to handle the cost of these? When you use Apple Intelligence, you have to connect it to a model, and that model usage is not free. But there is a wild maker attitude among some that reminds me of the early days of Linux back in 1993, when I downloaded 50 3.5″ floppies worth of Slackware 0.99.11 Alpha with the idea that this was a new wave of something big coming. I feel that way now, but with my Apple devices, there is no Apple AI beyond Apple Intelligence, which uses OpenAI behind the scenes. While Apple Intelligence is much better than the brain-dead Siri, there’s not much happening there. Given that tiny models (see (NanoClaw)[https://nanoclaw.dev]) could possibly run locally on your phone, what could the future look like for Apple devices hosting LLMs?
Apple Intelligence simplistically allows you to connect to a model, and that’s it. There is no plug-in system to allow for any kind of orchestration. The real power of OpenClaw lies in how it manages and leverages your personal email, notes, calendar, reminders, and contacts, knowledge that it gleans from reverse prompting and interviewing you, and how it integrates with other APIs and messaging services like Discord or Telegram. These are all sources that are readily accessible on your devices today. One of the first things I did with OpenClaw was set up a few projects to address immediate needs focused on summarization and aggregation. Getting a summary or distilled information is crucial when you’re overwhelmed by a flood of information. Are you subscribed to newsletters? If so, for those with similar interests, don’t most of the links tend to repeat the same content? There’s a lot of cognitive load in filtering through all that, and unsubscribing isn’t always an option, so it would be helpful to remove duplicates. Apple will need to allow for some kind of integration system that will allow orchestration, if it is going to a compete in the personal assistant arena.
Project 1: I set up a Kanban-style dashboard for my tasks. Throwing everything I need to do into it is crucial for getting an overall view. I use columns such as Recurring (openclaw crons), Backlog (where ideas start), To Do (openclaw knows to look here to begin working on an idea or story), In Progress (openclaw is working on it), Review (openclaw needs my attention on a task), Done (completed tasks), and Rejected (ideas that get a No). This is simply basic task categorization, and apps like Todoist have been doing it for a while. Sure, you could make your own Apple Reminders lists to match, along with some shortcuts kerjiggering it together, but seriously WTF, it’s just another way to view the same data. Nobody at Apple thinks a Kanban view is valuable?? Granted, having mine as a node process managed by launchd at 127.0.0.1 on my Mac and accessing it via Tailscale from my iPhone is a bigger kludge, so it’s far from ideal for a web view. Still, having OpenClaw manage it works quite well when you ask it to work with you through the To Do list.
Project 2: I created a daily newsletter of my own. Anyone who knows me knows I have a long-standing love of RSS going back to 2004 as far as my blog. This combines the items above with the output of RSS feeds managed by a modified version of rss-agent-cli, which also summarizes articles as one-liners using Gemini. I run it as a cron job to dump articles into .md files in an Obsidian Vault where I can view them. The newsletter pulls information from my Google resources and these RSS feeds, then generates three project ideas based on my interests and adds them to my Kanban Backlog. The OpenClaw agent comments on each section, making it feel more personalized. It’s a level of customization I would have expected from Apple.
Project 3: OpenClaw has an API key for my Home Assistant server. Nuff said there.
And honestly these are toy examples. Use of Opus to add features to existing open source projects, like in the example above for rss-agent-cli, helps further OpenClaw’s reach. Claude itself can do much of this today with its links to Google Mail, Google Drive, and Google Calendar, but while Anthropic is obviously taking cues from OpenClaw on integrations, the slow pace that they are making is almost as bad as Microsoft’s additions into Copilot. I understand that these developments prioritize safety, but it highlights the difference between the pace of open-source projects and that of commercial enterprises. Meanwhile, Apple’s efforts have taken much longer, almost as if they are afraid to take any step at all. Without functional integrations for Apple Intelligence in iOS 27, Apple risks looking truly ridiculous because delaying for another year will mean that they plan to enter AI in a very different landscape than where we are now. The AI landscape is shifting every month and any delay means you are left behind. If that’s the case, they will never catch the wave and will remain behind, sitting in the calm waters while the wave passes. Apple already has a rich intention library that they make use of in their shortcuts. The foundations are already there. Could it be that Apple is getting lost in how to handle context of the hot mess that most people’s Notes and doc folders are filled with?
What is the point of this? Some of you might say that it’s basically what you already have in your device’s notification center. No, that disorganized mess is the equivalent of the 10,000 unopened inbox messages most people have “as a system.” (LOL) What I mean is that with the overwhelm of all the apps and sources these days, where everything looks like an inbox, striving for Inbox Zero or even just trying to get your arms around it all is a challenge. Aggregating all that into one view that can distill things without manufacturing alarming headlines (if that’s even possible) can reduce cognitive load and surface the most important details—like alerting you weeks in advance about a birthday you might otherwise miss, efficiently adding items to a shopping list for an upcoming trip without duplicating what you already own, finding fast shipping options, sourcing items locally, bringing together your health data with your diet trends with your exercise sessions, browsing job descriptions based on what it knows about you, or keeping you updated on trending topics that are actually meaningful rather than getting lost in doomscrolling, or letting you know that you haven’t opened a certain newsletter for six months and should just unsubscribe. Starting documents or projects, even if the work is only begun or just partially completed correctly, is no different than using a template, and you, the human, step in to finish it up.
There will still be a set of people that focus solely on “AI as theft”, but I have already weighed in on that. There are also many that focus on AI as a tool that will kill us all. The truth is that, bubble or not, there will be no pause in AI research and progress, the giant data centers are going to come whether grids and resources are ready or not, and there is no stopping the giant companies that are blowing up valuation records – at least not for the next few years. The real question is to what degree they will race down the roads with this, and how many people will get run over for the sake of progress. And the question of UBI is a whole other thing to discuss.
Before Apple can be involved in this, unless they only plan to sell us hardware, they need to start by fixing dictation on their devices. There’s a reason there is a whole market of AI dictation apps. Apple’s dictation just sucks.
Playing with Sharp Objects. 