Quanta Bits - August 23, 2026

Hi everyone,

I spent the week testing Grok Bot, xAI's new always-on AI helper. It was much easier to start than OpenClaw, the self-hosted helper I tried. I did not have to set up and maintain another computer. Before I trust it with serious work, I need to understand what it can access, which models and computers handle each task, and what it really costs. Overall though, I'm very happy with it and have been using it more and more! Definitely I recommend a look.

Cheers.

Reza

More than a million people used LinkedIn's new "Seems like AI slop" report option in two weeks, and posts it classifies as low substance are getting about 40% fewer views. Anthropic says future Claude models will leave a detectable pattern in their word choices, showing probable Claude involvement but not who wrote or approved the work. US patent law still requires human inventors when AI helps design a drug. Together, these developments make AI involvement more visible while still holding a person responsible.

Harvey, a legal AI company, tested AI search across 9,288 fictional law-firm files. Two models answered correctly when they found the right files but covered only half the required facts. They stopped too early. Cerebras encountered a related problem as its internal collection grew: broad searches returned too much irrelevant material. Cerebras limited searches to the relevant project, converted Slack discussions into summaries, and added other search methods because vector search alone missed information. Both cases show that access to more information does not guarantee a complete answer. Relevant evidence can still be missed unless search is narrowed and designed to keep looking.

Cursor's Origin can store code, manage proposed changes, and mirror GitHub projects while GitHub remains their official source. Slack Code brings coding assistants into shared channels where teams can see plans, changes, previews, and discussion. CIOs now have to decide who can change code, how changes are approved, where the official version lives, and what record is kept.

OpenClaw helped popularize a self-hosted, messaging-first personal AI agent earlier this year. Peter Steinberger's open-source project ran on your computer, connected AI to local files and applications, remembered prior work, and answered through Telegram, WhatsApp, and other messaging apps. It made the idea of an always-on helper tangible.

In February, I installed OpenClaw on a dedicated Mac Mini and wrote about the experience. Yep, I was part of the Mac Mini craze! Spock, my agent, ran scheduled chores and stayed available through Telegram. I also became its system administrator. Updates sometimes left the setup out of sync, connections failed, and I managed its permissions, memory, model choices, security, and computer. At one point, I used OpenAI's Codex coding agent to repair it after OpenClaw couldn't reliably repair itself.

An AI model (think Opus, GPT, Grok) is only one part of a working agent. Around it sits an agent harness, which prepares what the AI sees, remembers prior work, gives it tools, and manages each step. The broader operating layer adds the computer or cloud environment, permissions, schedules, approvals, logs, and the interface you use. Together, they determine where work runs, what the agent can touch, and what you can review afterward.

For me, OpenClaw was the first time I assembled most of that layer myself. Managed products had already packaged parts of it, and several have expanded their remote options this year. Anthropic's Remote Control and Dispatch let users continue or send Claude work from another device, although I found the split confusing and performance inconsistent. Anthropic updated Remote Control earlier this month. I have been testing the newer version, and so far it works much better. OpenAI's Codex worked better from my phone but remained awkward. Nous Research's open-source Hermes offered more model choice and required more setup, but outside Telegram and similar interfaces, still no dedicated mobile app.

In short, if you wanted to have a multi-agent experience, something available to you 24/7 from different devices, available tools had flaws and limitations.

On August 11, xAI, the company behind the Grok models, released Grok Bot in early beta. Each Bot is a named AI helper with a job, its own conversation and memory, and access to tools. My Bots share a managed cloud computer and remain available after I close my laptop. I can continue the same conversation from my phone.

Setup is far easier than OpenClaw or Hermes. I didn't have to buy and configure another computer or maintain the software that kept it online. Everything works out of the box. I started with one Chief of Staff and added Bots when real jobs appeared. Each received a short job description. Those instructions define the work, while permissions and approval rules control what the Bot may do.

I use the Bots for several recurring jobs. The finance Bot helps me review Quanta's finances and spending. The CTO organizes my technology research and what I am learning. Book Advisor maintains my reading list and book notes in Obsidian, the notes app I use. This week, it gathered my notes after I finished Other Minds, asked Codex to write the final entry, and saved it.

Grok Bot can work with files on my computers as well. I can start a request on my phone and have a Bot update notes on an always-on Mac. I can also connect tools from OpenAI or Anthropic through APIs or command-line software. I use Codex and Claude Fable this way, but it requires technical setup and specific instructions. The handoffs aren't automatic.

One genuinely useful feature is the ability to connect more than one account from the same service. I connected both my personal and work Google accounts under the same Grok Bot account. I have not found another product that makes this equally simple.

That convenience comes with a tradeoff. All of my Bots share the same cloud computer, including its files, browser sessions, and logins. I would connect only accounts that I am comfortable making available to every Bot.

Some basic pieces still fail. The Zoom connection does not work in my setup, and I have restarted the virtual computers several times after they became stuck. This is still beta software.

When a Bot's output is weak, I can't easily diagnose why. In my individual account, I can't see which AI model answered or how much reasoning it used. xAI says team usage reports show which model served each request, including when the system switched models, but users and administrators cannot choose the model. I also cannot adjust the reasoning level or trade speed for deeper analysis as I can with tools such as Codex.

My computers hold different files, applications, and account access. I can name one in my request in Grok Bot, but I can't guarantee that the Bot will use it or confirm afterward which computer ran the work. That leaves me guessing whether a weak result came from the AI, my instructions, or the surrounding system.

Grok Bot can be expensive too. I had to get Cursor Ultra, the $200-a-month plan, for ongoing access. Cursor has since expanded access to cheaper plans, although I have not tested how far their allowances go. After less than a week, I have used 85% of my weekly allowance with two days left, without doing any coding. Usage may become a real constraint.

Of the products I have tried, Grok Bot is the easiest starting point for someone on the fence about an always-on helper. It hides much of the setup OpenClaw or Hermes users had to manage, which helps explain the attention it has received.

The tradeoff is control and cost. The tool is opinionated, and sometimes that is the right choice for simplicity. More specialized work usually requires more control over the model, reasoning level, and how work is routed. I would recommend trying Grok Bot, specially if running the infrastructure has kept you from using an always-on helper and you are comfortable with beta software. I have been using it more and more this week, at the expense of my other agents. My development and strategy work, however, still stays with Claude and Codex.

Source: Douglas Boneparth on X

Moderna and Merck said their personalized mRNA treatment for high-risk melanoma met both goals in a late-stage trial. After surgery, patients received a treatment based on mutations in their tumor plus Keytruda, a cancer immunotherapy drug. Compared with Keytruda alone, the combination delayed the cancer's return and its spread to other organs. The companies have not released detailed results, so the size of the benefit is unknown. Moderna calls this the first successful Phase 3 trial of a personalized mRNA cancer treatment.

Reasoning level: A setting controlling how much time and computing an AI model uses before answering. Higher settings can help with difficult work but take longer and may cost more. Some tools let users choose; Grok Bot does not.

Orson Scott | 324 pages | ★★★★☆

I watched the Ender's Game movie a while ago and only vaguely remembered it. I recently decided to read the book. At the same time, I have been reading The Infinity Machine, which notes that Ender's Game made a deep impression on DeepMind co-founder Demis Hassabis when he was young.

The military's use of children as soldiers was uncomfortable to follow. The simulations and endless practice shaping Ender's decisions also made me think about how we train AI models. If repetition works on machines, why wouldn't it work on humans? Or did I get that backwards?

I especially liked the ending. Some Battle School sequences felt repetitive, and the action was occasionally hard to follow in the audiobook. I cared more about the manipulation, politics, and moral machinery than the games. Still, it is fast, imaginative, and much more unsettling than its military-adventure setup suggests.

Prefer to listen? Quanta Bits is also available on Apple Podcasts and Spotify.

*I collaborate with Spock, my AI agent. He researches extensively: scanning, filtering, and surfacing what is relevant across my business. I read, listen, and watch what resonates, and decide what matters. I provide the direction, and we draft together. The editorial judgment is mine. He would tell you the same. Most logical. 🖖