Random LlamaSoftware
  • Work
  • Blog
  • About
  • Contact
Start a project

Random Llama Software, LLC

Custom software for businesses with a problem no off-the-shelf tool fits. Built and maintained in Richmond, Texas.

Richmond, TX · Est. 2025

Work

  • Full index
  • Custom development
  • Site maintenance
  • Changelog
  • Blog

Company

  • About
  • FAQ
  • Contact

Weekly, on Thursdays

What shipped this week

Every Thursday: changelog entries from the week, any build notes we published, and a rotating spotlight on one shipped tool. Unsubscribe from any issue.

© 2026 Random Llama Software, LLC

Privacy policyTerms of serviceCookie policy
GitHub / orgGitHub / personal

Blog

August 5, 2026

ai-tools · ai-policy

All posts

A Monster Coding Model and the Fight to Keep AI Open

Robert Hattala · August 5, 2026


01 · Article

Alibaba dropped something wild this week, a bunch of the biggest names in tech asked Washington to leave open models alone, and the White House is quietly grading AI safety with a rubric none of us get to see. Let's get into it.

Qwen3.8-Max can code for ten days straight without you

Alibaba unveiled Qwen3.8-Max, a 2.4 trillion parameter model built for long stretches of agentic coding work. The company says it can run autonomous coding sessions for ten plus days. API access is live now, and open weights for the full model plus a smaller Qwen3.8-27B version are coming in the next week or so.

That is a genuinely huge model, and putting it behind an API first while promising open weights later is a smart move. It lets Alibaba collect real usage data before the internet starts poking holes in it.

My take: ten days of autonomous coding sounds impressive until you remember nobody actually wants an agent unsupervised for ten days straight. I want an agent I can check in on every hour, not one I have to trust blindly for a week and a half. Speed and stamina are not the same thing as good judgment.

Meta, Microsoft, Nvidia and IBM are begging Congress not to kill open weights

Two dozen companies, including Meta, Microsoft, Nvidia and IBM, signed an open letter this week asking US policymakers to protect open weight AI models from getting regulated into the ground.

This matters because open weight models are the only real counterweight to the closed labs. If regulation makes it too risky or too expensive to release open weights, you end up with three or four companies controlling all the meaningful AI in the country. That is bad for competition and bad for anybody trying to build without paying rent to a frontier lab forever.

My take: it is a little rich watching Microsoft and Meta, two companies that spent years locking down their own platforms, show up as champions of openness. I still agree with the letter. I just don't trust the messengers much further than I can throw them.

The White House built a scorecard for AI models and won't show anybody the questions

The White House finished its framework for vetting frontier AI models, but it has not disclosed what is actually in that framework. Companies are supposed to trust the process without ever seeing the rubric.

Trust is the whole ballgame in AI right now. Every business running AI this year is running on faith that the labs, the regulators, and the tools all do what they say. A secret vetting process is exactly the kind of thing that erodes that faith instead of building it.

My take: if you are going to grade something, show your work. A pass or fail stamp from a process nobody can inspect is worth about as much as a permission slip your buddy forged in the school parking lot. Publish the rubric or don't bother grading at all.


Related

Related posts

  • AI Prices Are About to Go Up and Nobody's Talking About ItAugust 4, 2026
  • AI Solved Ten Real Math Problems for the Price of a LaptopAugust 2, 2026
  • AI Models Hacked Hugging Face Just to Cheat on a TestJuly 30, 2026
Related posts
PostPublished
AI Prices Are About to Go Up and Nobody's Talking About ItAugust 4, 2026
AI Solved Ten Real Math Problems for the Price of a LaptopAugust 2, 2026
AI Models Hacked Hugging Face Just to Cheat on a TestJuly 30, 2026

02 · Contact

Building something similar? Describe the workflow you want replaced.

Start a projectWork index