Quanta Bits - February 14, 2026

Hi everyone,

Big week. I wrote about OpenClaw last week. This week, I actually set it up and lived with it. A snowstorm on Saturday gave me a good excuse to stay inside and give it a real try. The promise is that an always-on agent makes you more productive. My experience after a week? You end up getting sucked into a lot more things than before. I'm exhausted after a week of working with my new agent, Spock.

(PS: You'll notice in the image that Nano-Banana had a hard time getting Spock's "live long and prosper" right. Such is life working with AI.)

Last week I wrote about what OpenClaw is and why it matters. This week, the personal version. I spent 7-8 hours setting up my own always-on AI agent on a dedicated Mac Mini, and that was just the start. Another day of tweaking to get the base configuration right. The estimates I'd seen ranged from 20 minutes to an hour. The reality with raw open-source software is significantly more, even with a technical background.

The agent's name is Spock. A good friend asked why not Data? Well, Spock is flesh and blood. Had to pick the side of some measure of living in this world. In reality, my Spock weighs more heavily towards its human side of irrationality and "assuming things" than the logical Vulcan. It takes a lot of patience even now to get it to be more predictable than I expected.

It runs OpenClaw on a dedicated Mac Mini, connected to Telegram so I can message it from anywhere. It has access to my Obsidian knowledge vault, can search the web, and manages my tasks. I dictate to it using voice-to-text, and it responds with generally thoughtful, context-aware answers.

Here's what I've learned so far about living with an autonomous agent:

The technology works. Within 12 hours I went from "how would I use this" to "what else can I give it." It's genuinely useful once running, managing my tasks, reminders, and notes. But "works" and "production-ready" are different things. Every lesson above, security, costs, boundaries, memory, is a governance question that enterprises will face at scale with autonomous agents. Better to learn them on a Mac Mini first.

Full blog post coming soon with the complete setup guide, security architecture, and lessons learned.

The Economist ran a piece last November about the AI companion industry. One line stuck with me: "If everyone has a virtual friend in their pocket that is always ready to take their side, this could create 'personal echo chambers of validation' in which they are not challenged to think for themselves."

The article was about chatbot romantic partners and AI therapists. But the observation applies to something closer to my daily work.

Now I run a 24/7 AI agent at home. It handles note-taking, research, task management, even curating what I should work on next. The friction it removes is real. I used to spend 10-15 minutes after a meeting writing up notes and pulling out action items. Now it happens in seconds. I used to sit with an article for a while before deciding what to do with it. Now it gets clipped, summarized, tagged, and slotted into a draft pipeline before I've finished my coffee. Mind you, I still read the articles!

Here's what I've started noticing: some of that old friction was doing something useful. The 15 minutes of note-taking forced me to replay the conversation in my head. The act of manually summarizing an article made me decide what I actually thought about it, not just what it said. The friction wasn't wasted time. It was thinking time wearing a disguise.

What's replaced it is a different kind of treadmill. The bot produces research reports overnight. It suggests tasks. It organizes my reading list and tells me what's "up next." I wake up to a backlog of AI-generated output that I now need to process. The bottleneck shifted from creation to evaluation, but the volume of things demanding evaluation keeps growing. I'm not doing less work. I'm doing different work, faster, definitely more of it.

Advait Sarkar, a researcher at Microsoft, put it perfectly in a recent TED talk: "We've solved the problem of having to think. Unfortunately, thinking wasn't actually a problem." His team's research shows that AI-assisted workers produce a narrower range of ideas, apply less critical thinking, and remember less of what they worked on. Even mundane tasks, the kind we're most eager to automate, serve as essential cognitive exercise. Skip them long enough and you lose the muscle for the hard stuff. As Sarkar puts it, we've become "middle managers for our own thoughts."

The convenience is so good that it's easy to stop noticing what you've outsourced. Not just tasks, but the cognitive work that used to happen while doing those tasks. The question isn't whether to use these tools. The question is which friction to preserve on purpose.

Sources:

My other 24/7 agent, creating more work for us rather than saving, but with so much more emotional returns 😊

OpenAI starts running ads in ChatGPT
OpenAI flipped the switch on advertising for its Free and Go subscription tiers in the US. The company claims ads are clearly labeled, optimized for helpfulness, and won't influence ChatGPT's responses. Worth comparing to common criticism on social media: "Actions speak louder than words. If AGI is around the corner, why would you bother with ads?" Google has no plans to add ads to Gemini. Anthropic expects to break even by 2028 without ads. For enterprise buyers, the question is whether your AI vendor's business model aligns with your interests or with advertisers.

AI's Generation Problem: 200 Answers, 1 That Matters
Google DeepMind pointed an internal Gemini model at 700 unsolved math problems left behind by Paul Erdos. The AI generated 200 candidate solutions. After expert review, 63 were technically correct. Of those, only 13 addressed the actual intent of the problems. And of those 13, only 2 were genuinely novel, and only 1 was actually interesting. The funnel tells you everything: 200 to 63 to 13 to 2 to 1. For organizations deploying AI in knowledge-intensive workflows, the implication is practical: budget for evaluation time, not just generation time.

AI's Super Bowl moment looks familiar, and that's worth noticing

AI companies dominated Super Bowl Sunday ads this year. Anthropic, OpenAI, Google, Amazon, and even Svedka ran AI-focused spots, with 30-second slots costing up to $10 million. The most headline-grabbing move: Crypto.com CEO Kris Marszalek purchased AI.com for a reported $70 million and launched a personal AI agent platform with a Super Bowl ad. The comparison that matters is 2022, when FTX, Crypto.com, Coinbase, and eToro all bought Super Bowl ads. Nine months later, FTX collapsed. The same CEO who ran that Crypto.com ad is now running AI.com. But unlike crypto in 2022, AI has substantial enterprise revenue and real use cases. The question isn't whether AI is a bubble. It's whether the marketing is outpacing the maturity. (AI.com crashed shortly after the ad aired, overwhelmed by signups.).

Anthropic owns 80% of enterprise API spending
Ramp data from January 2026 shows Anthropic dominated nearly 80% of API spending, the kind that happens when companies integrate AI into their own products and workflows. Meanwhile, OpenAI's ChatGPT had nearly 900 million monthly active mobile users. The enterprise vs. consumer split is now quantifiable. Anthropic wins where companies build on top of AI. OpenAI wins where individuals use AI directly. For CIOs, this is a concentration risk question: if 80% of your organization's API spend flows to one provider, your AI strategy has a single point of failure.

7% of AI agent skills contain critical security flaws
OpenClaw partnered with VirusTotal after roughly 7% of listings on ClawHub, its skill marketplace, were found to contain critical security vulnerabilities. Every published skill is now scanned using VirusTotal's AI-powered Code Insight tool. The 7% number is worth sitting with. That's roughly 1 in 14 skills containing flaws serious enough to be classified as critical. For CIOs, this is the agent sprawl governance problem made concrete: when anyone can publish a skill and anyone can install it, you've recreated the shadow IT problem with a wider attack surface.

Every week, a small group of practitioners and technology enthusiasts meets to share learnings about automation, AI, and emerging trends. We call it Quanta Lab. Invitation only. Here's what we've been discussing lately.

Peter Steinberger (creator of OpenClaw, formerly PSPDFKit) described his daily workflow in a recent interview: 5-10 parallel AI agents running simultaneously, each "cooking" on a different feature for 20-60 minutes while he switches between them. He compared it to a chess grandmaster playing multiple boards at once.

The perception that AI makes development easier misses what's actually happening. The cognitive load doesn't disappear. It shifts upward. Instead of writing code, you're managing parallel workstreams, making constant architectural decisions, and context-switching between features at a pace that didn't exist before. Steinberger said he's "never worked harder" than he does now, not because he has to, but because the speed of iteration makes it addictive and mentally consuming.

As I mentioned at the beginning, that's been my experience. There's a rush you get in knowing an agent is always working while you sleep, thinking, managing, organizing, researching. But there's always more to give it, and the boundary between work and personal time gets harder to hold. Sometimes I question if the agent is working for me or I'm working for the agent.

Sources:

The Vercel skills CLI (npx skills) lets you search and install Claude Code skills from a global repository. One of our practitioners tested image generation skills that use Google's Gemini model, walking through style, layout, texture, mood, and typography choices in the CLI, then producing custom blog illustrations. The quality was impressive for what is effectively a collection of text files.

The broader observation from our lab: "The world's best knowledge is quickly being abstracted into downloadable skills. Everyone gets access to the best principles." Governance skills, security skills, content skills, all discoverable and installable in seconds.

This is exciting and intimidating at the same time. If everyone has access to the same world-class principles, differentiation shifts from what you know to how well you apply it. For consultants, this means the value isn't in having the framework. It's in knowing which framework fits which situation, and having the judgment to adapt it.

Pierce Brown
★★★★☆

I finished the first Red Rising book in May 2025, gave it 2 stars, and wrote "Won't read the other parts of the series." It felt too much like Hunger Games, all war games and not enough plot.

Five months later, I gave Golden Son a shot. Within a week I understood why people love this series. Minor spoilers: the ending hit like the Red Wedding in Game of Thrones and I immediately started the next book. From that point, I read five books back-to-back. Nine months and six books later, here I am, patiently waiting for Red God, the final chapter that's supposed to come out soon. A big commitment, over 3,000 pages across the series (all on audio), but it was worth it.

I really liked the action. The books are light on technology and focused more on story and momentum. The action scenes are cinematic. Definitely worth a try, but beware of the first book sounding too familiar!