One day of usage on the Claude Code Max 20x package. All legit use, too. Primarily running dynamic workflows to handle multiple large, complex tasks in parallel. I guess I'll soon need a package for each day of the week. Not sure whether to laugh or cry 😅

Not that anyone asked, but there are 4,701,128 publicly indexed podcasts, totaling 148,838,924 episodes. That takes roughly 793 GiB of space just for metadata. Plus ~16 TB for transcripts. ~470 GB for artwork. And god knows how much for the audio and video.

I don’t mind being scraped. But show some respect and pace yourselves while devouring all that yummy data. These servers weren’t built for this kind of load.

Codex is much better with 5.6 Sol, but still well behind Claude Code. It drifts off target, forgets what it’s doing, breaks the code it’s working on, needs very specific instructions, and runs slow. Rooting for it to catch up. We need the competition. PS, there was no screenshot😅

Now that all my Claude accounts are maxed out, I'm using Codex for now. I didn't realize just how slooowwwwwww it gets the moment a task gets even mildly complicated. Devs using Codex must have a lot of free time on their hands while waiting 😄

The biggest problem I haven’t figured out yet with Claude Design is getting Claude Code to actually implement it. It always just makes a sloppy, similarly colored version instead 😕

I’ve already cancelled ~1000€ in annual app and plugin subscriptions. I just have AI build them for me now. Today I cancelled two plugins that synced billing data to my accounting. The new custom plugin took over the state from the old ones and resumed without missing a beat.

I’ve noticed that with Opus 5, Cloud Code asks more questions mid-run, but instead of surfacing them or waiting for a response, it just plows on with “since you didn’t respond …”

Pancake Day at GitHub?

I need to stop using Ultracode in Claude Code. 56% of my Max 20x package credits are gone in just one day 😬

These little menu items dedicated to agents are becoming the norm in most software

It would take only about eight years to pay for itself. That's assuming 100% utilization 24/7 of six fully maxed Claude 20× Opus accounts, and absolutely zero hardware failure or obsolescence.

I like the Claude voice. But I just don’t have the patience for it. Combining the locked 1x speed with Claude's tendency to reply in miles of text for everything just takes forever.

Probably not a popular take, but after finishing The Innovators by Walter Isaacson, I’m even more convinced that everyone who can should work from the office. Almost nothing in that book came from a lone genius.

Lately, Claude Code has started ending sessions before completing the task, saying it’s about to run out of context.

Maxing out the limits on 3 Claude Max (20x) accounts and 3 Codex Pro (20x) accounts in one week was way too simple 🫪

OpenAI's proprietary GPT-5.6 Luna is on par with the open-weights Z.ai GLM-5.2 on the AI Intelligence Index. Yet Luna costs $1/$6 vs. just $0.77/$2.42 for GLM.

Most company automations tend to be forgotten and left to do their job for years without anyone revisiting them. When you consider this in light of the cost of every token they crunch through over those years, the difference can end up being more than a company's entire IT budget.

ai•Lab@ailab.ee · 2w ago

Don’t just blindly automate your company’s work with whatever LLM you’re used to. This graph is a perfect illustration of why it’s important to pick the right model for every task. The cost difference over the years will be huge.

OpenAI app is such a mess. I have no clue where I’m even supposed to start a coding tasks on mobile anymore. And how do I just start a regular chat on desktop? 🤷‍♂️

When Fable 5 and GPT-5.6 Sol refused to audit my code + infra security, I tried the open-weights GLM-5.2. It actually did much better on infra than Opus 4.8, though Opus was still better at code. Not sure if it was a fluke. Will need to test it more.

I guess it’s not gonna be GPT-5.6 Sol either. Codex is somehow even worse at this than Claude. After 43 minutes of burning tokens across 9 subagents in Ultra mode, it just told me I can’t see the results 🤦‍♂️

I guess I'll try using GPT-5.6 Sol for security work now

I’m scared … the amount of subagents Claude Code fired up in Ultracode 🫪

Ultracode selection in Claude Code mobile app is fun 😄

I noticed today that 5.6 Sol is the first OpenAI model I'm actually using for my daily work. It doesn't hurt that the limits for the same monthly fee seem much more generous. I haven't hit them once yet, even while running long tasks in Ultra mode.

To protect your sensitive data from agentic attacks, start adding prompt injections and references to Tank Man from the 1989 Tiananmen Square massacre. https://arstechnica.com/security/2026/07/now-defenders-are-embracing-the-prompt-injection-too/

Now, defenders are embracing the prompt injection, too

Context bombing" tricks hacking agents into shutting down before they can do harm.

arstechnica.com

One strong benefit GPT 5.6 has over Claude in design is its ability to understand, use, and manipulate images.

So disappointing to see Anthropic and OpenAI trying to kill potential competition in the open-weights LLM market and pull up the ladder behind them 🙁 https://www.bloomberg.com/news/articles/2026-07-13/anthropic-openai-warnings-prompt-distillation-debate-in-dc

Washington Is Looking to Keep China From Training Its AI on US Models

Anthropic and OpenAI’s warnings about “adversarial distillation” are reinvigorating one of the oldest debates in Silicon Valley.

bloomberg.com

(1/2) If Apple can actually pull this off, then we’re all moving most of our daily workloads to on-device LLMs much sooner than we thought 🤞