Just realized my Pixel 11 Pro promo for doesn't allow Gemini 4 Argon. FYI, in case anyone has it. It's an otherwise great promo: 24 mo discounted Google Fi, 6 mo free Google One (5TB), YouTube Premium. I'm actually pretty happy with the Pixel promo; win some, lose some.

What are AI tokens? I like to think of them as bytes. When people say: 1 million tokens, I loosely translate it to Megabyte. Billion ~ gigabyte. Tokens are still just 1s/0s. Not a literal translation, but also not far off and helpful to demystify data. x.com/0xAndrewShu...

We're adding the Command Firewall to Branchmore to block unsafe tool calls. Our engine identifies your agents' risky tool calls, then make it easy to add firewall rules. Branchmore identifies risky tool calls ➡️ suggests firewall rules you can add ➡️ delivers them to pre tool use hooks for blocking.

Heard a talk by an OpenAI engineer talking about his 150B/mo workflow. One interesting tidbit was that disk space was a problem, bc of how many parallel agents he was running. They each needed their own workspace. Another fascinating bottleneck as we scale AI usage.

I hear about GPU bottlenecks + RAM prices but CPU/disk are also in short supply. GPUs may do "thinking" (inference) but CPU's power the "tool calls" (code/APIs). Fascinating all the places bottlenecks are appearing. spectrum.ieee.org/ai-cpu-come...

Roll Over GPUs, CPUs Are Back In the Spotlight

Agentic AI has made CPUs the new performance bottleneck

spectrum.ieee.org

1/ Is Chinese more efficient than English for token use? Turns out: no! Or more broadly: how do other languages compare, like Arabic, Spanish, etc? There was a paper Jun 2025 that claimed English was less efficient, but that was debunked Apr 2026. Papers in thread. 🧵

Two useful heuristics when you estimate how many tokens you might use: "1 token is ~4 characters", and "3 English words are ~4 tokens." Two common context window limits are 200k and 1M. Avg English word is 5 characters, so that's ~10k words and 50k words, respectively.

Another PSA: if you use AI with API keys, look for spend limits in the configurations. This is what OpenAI's API platform looks like. In the past week, we've seen breaches and key leaks leading to huge ($90k) bills! It's an easy change to limit a huge surprise bill; easy to change right now.

Enjoying spending more time in OpenCode. Logged in w OpenAI for Sol to focus on the UX. A few initial thoughts: I like the code diff view when edits are made and the sidebar with the session title, tasks, context usage. I'm curious abt OpenCode advanced features. Any recs?

Oof, Opencode breached yesterday. Seems to be the Console (web) users, not folks who only run OpenCode locally. AI provider API keys (OAI, Ant, Google) keys leaked for 1.4% users, and folks are reporting their AI credits were consumed. x.com/opencode/st...

What features would you recommend I try on OpenCode? So far, using Sol (not Go). I've been enjoying the interface so far. I like the code diffs when I want to drill into a particular task the agent's doing. The always-on sidebar is also nice TUI for ctx window usage, tasks, etc.

I've enjoyed starting a blog this year, but it's been terrible for traffic. I don't remember it being this bad 10 years ago. Blogging and SEO for sales for a solopreneur starting out in 2026. Given search traffic dropped 60% for small business, the juice isn't worth the squeeze.

Google Search traffic has dropped 40% for publishers over the last 2 years, which impacts businesses trying to sell on Google. There seems to be a spectrum - small publishers tended to get hit harder at ~60%, whereas larger publishers tended get hit ~20%. Brutal.

Blogging is dead in 2026. Been talking w/ friends abt their small business blogs; traffic down 40-60% in the last 1.5 yrs bc Google AI Overview. Sales drops 50% overnight. This has always happened with Google, but this time feels permanent. Anyone encounter something similar?

It's been a rough few weeks for AI. This week: AI datacenter Coreweave stock/bonds having trouble, Character.ai filing for bankruptcy. A few weeks ago: Character.ai CEO left to be Disney's first CTO.

Google and Microsoft making moves with AI models. Gemini 4 Argon compares to GPT-6 Sol. And Microsoft's MAI-Base-2 model is comparable to OpenAI's Sol but cost 25% pretraining compute. I hear Gemini 4 has been good communicator; effective and concise. Looking forward to trying.

Just realized that resuming an old agent session triggers a bunch of cache write tokens on the old transcript. So: I'm gonna avoid --resume, especially big sessions.

1/ PSA: this is an example of how agents can run unsafe tool calls and misuse environment files with keys/tokens. A leaked key can quickly rack up a big bill or data leak. In this case, a $90k bill.. yikes! Glad GCP cancelled the charge, but that's a close call.

Been thinking about content marketing, but the sharp drop in SEO traffic in 2025-2026 for small blogs has been brutal. When I talk with friends who run blogs or small businesses, content marketing seems to have shifted to social media. Anyone else have Google traffic drop in 2025-2026?

"Try new tools" is great advice in thisd AI environment, but there are so many tools! On the one hand, I learn more quickly. But on the other hand, learning new tools eats into other work. I've been improving my discipline in making experiments lightweight: 🧵

I just realized that Grok has a comparative advantage among agents bc it can search X content. I've found X to be useful for technical and business insights in a unique way compared to Reddit, LinkedIn, Google Search, BlueSky. I checked GPT/Gemini/etc and they can't read tweets. (Both X/BlueSky).

Is it useful to build in public these days? I've read commentary this isn't useful / interesting anymore. Open to advice!

Try Haiku 5.5 to reduce token costs: 20x cheaper than Sonnet, 40x vs Opus. It's Anthropic's answer to Lunamaxxing: using Luna on Max reasoning. With the release of Haiku 5.5 last week, seems intelligence notably increased + costs reduced. What's been your experience w/ Haiku?

🙏 Thanks for everyone's support with my Product Hunt posting yesterday! Now it's back to digging into our agents' logs for productivity insights. ⛏️ Hope everyone has a great weekend.

0xAndrewShu@ashu.co · 2w ago

☀️ I just launched Branchmore on @ProductHunt! Branchmore tracks tokens and agent traces to find problems with token costs, unsafe commands and unproductive agents. Then it converts those learnings into fixes for your AGENTS[.]md and Agent Skills. www.producthunt.com/products/bra...

☀️ I just launched Branchmore on @ProductHunt! Branchmore tracks tokens and agent traces to find problems with token costs, unsafe commands and unproductive agents. Then it converts those learnings into fixes for your AGENTS[.]md and Agent Skills. www.producthunt.com/products/bra...

Branchmore: Track tokens and analyze sessions to find/fix agent problems | Product Hunt

Branchmore improves productivity for BOTH agents and the people who use them. To start, it combines token tracking (like ccusage) and session recording into one observability tool. Then our analysis g...

producthunt.com

Beyond Jev, what do folks think about GLiNER2.5 by Fastino. It's also a lightweight, encoder-only, open-source model. Similar inference pricing (cheaper, I think). But GLiNER can also do text extraction to JSON, not just finite decision outputs. Model: huggingface.co/fastino/glin...

fastino/gliner2.5-base-v1 · Hugging Face

We’re on a journey to advance and democratize artificial intelligence through open source and open science.

huggingface.co

Tried Jev in some judging/scoring/triage use cases. Drop-in replacement, but faster/cheaper. I could see it helpful with labelling higher data volume use cases to prioritize human effort, e.g. cybersecurity or other alerting (eng oncall, online marketing, support) workflows.

I'm planning on posting @branchmore.bsky.social on Product Hunt next Friday or Saturday. Does anyone have any advice for my first launch?

Useful prompting trick: repeating the prompt often boosts accuracy (for LLMs not using reasoning). Since LLM attention looks at previous tokens, the standard single prompt lets LLM "understand" each word only from the words already stated earlier in the prompt. (Interesting 2025 Google research)

Had a great time at @wearedevelopers.bsky.social in San Jose last week. Great pickup conversations with passionate engineers working on interesting things. Learned a few things in useful technical talks from engineers and teams I admire. This is a conference I think I'd enjoy coming back to.