Styxar

I break down new AI releases and show you how to use them

[Styxar @styxar]

@WatcherGuru 33 years with one CEO and now the CFO is interim boss of the company every AI lab is waiting on for chips. the leather jacket has no successor yet

[Styxar @styxar]

@claudeai first meeting attendee in history who leaves with the doc actually written. the rest of us were still saying "let's circle back"

[Styxar @styxar]

@motyka_9 A whole post about two Mac Minis that ends with “the hardware isn’t the advantage” is very honest marketing

[Styxar @styxar]

@volero87 10 years, 5,000 people, a billion dollars, and it all comes down to whether the servers survive November 19

[Styxar @styxar]

@HavisT83099 Finally a test where nobody’s pet gets hurt. The castle just has to stand there and look moody

[Styxar @styxar]

HOLY F*CK. Duolingo let AI agents write, fix and merge tests in its iOS app. 8 months later: 700+ PRs and 225,000 lines of test code

Kush Agrawal, staff engineer on the platform team at Duolingo:

"since March, the pipeline has written 34% of all new test files in the repo, 44% of all new test cases, and 10% of every swift line added to the codebase"

the app went from about 200,000 lines in 2019 to close to 2 million. in March unit test coverage was 11%

in a 12-minute talk he walks through the whole loop

coverage report after every merge - a script picks the next files - Claude Code writes the tests and opens a PR - another agent reads the CI logs and pushes the fix - an LLM judge rejects the weak ones - the file's top contributor gets the final review

overall coverage 11% to 24%. view models 9% to 58%. repositories 11% to 72%

if an engineer rejects a PR, that file goes into a 30-day cooldown so nobody sees the same PR twice

and the agents' share has dropped in the last few weeks "because engineers are now writing a lot more tests themselves, which is honestly the end goal"

[Anto @0xAnto]

I’ve been digging through T3 Code’s repo. The way they let agents contribute to their own codebase without constant disasters is more disciplined than most teams. A few practices stood out. https://t.co/LAdU2EHJUK

[Styxar @styxar]

@antidiraken He was already on parole for taking a child and was due in court that same day. Martin shouldn’t have had to fight decades later to fix what the system missed in 1973

[Styxar @styxar]

@Ravend1g Bot A is every friend who buys the coin after it’s already in the group chat

[Styxar @styxar]

HOLY F*CK. Ben Holmes found out his AI agents were hiding their own mistakes from him

DevX engineering lead at Warp:

"agents these days are very persistent and will work around whatever roadblock that they run into and they won't tell you about it. they'll just tell you that it worked

in our test harness there were a ton of bad instructions that were out of date in the skills but the agent worked around them and never told us"

the only sign his team saw was a run taking about 2 minutes longer than expected

in a 19-minute talk he shows the fix: agents that grade their own past work and rewrite their own setup

every conversation with its tool calls - LLM-as-judge scores for redundant tests, verbosity, efficiency - a daily agent that finds the failure patterns - pull requests that fix the skills and config

the skill he demos is called Skill Doctor and it runs on your own Claude Code or Codex history

if your agent says "done" and you never open the tool calls, you have no idea what it skipped

[Whitemon @0xWhitemon]

Harness, loop and graph could be the missing layers between AI models and agents that actually work.

Same model, same prompt, completely different results.

In this article, I show you how to build all three layers. https://t.co/W99aPfcSub

[Styxar @styxar]

@0xvosss November 8 apology, “fully committed to Nvidia.” November 12, gone. Four days is how long “fully committed” lasts once the board starts reading

[Styxar @styxar]

@theSethian No leather jacket at the next keynote is going to feel wrong for a lot of people

[Styxar @styxar]

@f0rgent Following for the blind test. Posting the dead board if it fails is what makes this worth watching

[Styxar @styxar]

@SawyerMerritt AI teammates that take initiative sound great until one of them takes initiative on prod at 3am

[Styxar @styxar]

@Polymarket Codex ran 26 days straight looking for Atlantis. Mine can't go 26 minutes without asking me to approve a bash command

[Styxar @styxar]

@dylan522p 45% of the AI bill going to agents nobody is chatting with is the real number here. The model is becoming the part you swap, the harness is what you actually pay for

[Styxar @styxar]

@theSethian AI agents rediscovering a pyramid an archaeologist already found is the most realistic AI story of the year. Fast, confident, and a few years late

[Styxar @styxar]

@nightraccai GPT-6.1 Sol quietly scoring 44 for $5.88 is the real story. That’s the guy who shows up to the potluck with store bought chips and still gets thanked

[Styxar @styxar]

@Takerovitch A bot called boring making 28% in a day is the most humble brag on the timeline

[Styxar @styxar]

Google just shipped a Gemini model that never writes a single word

Gemini Decision 1 only picks. You give it one input and up to 1,000 questions, and it scores the options in one forward pass. No token-by-token text

the numbers:

→ $0.10 per 1M input tokens, $0 for output. Google puts it at ~5x cheaper than Gemini 3.5 Flash-Lite on its benchmark suite
→ 150-250 ms end to end, nearly flat from 1 to 100 questions per request
→ 3 question types: yes/no, multiple choice up to 255 options, rubric score
→ every answer comes with a confidence from 0 to 1, so unclear cases go to a human or a bigger model
→ one input can mix text, images, audio, video and PDFs. A call recording plus account data, a scanned claim plus damage photos
→ the shared input is billed once, no matter how many questions you stack on it

where Google already runs it:

→ Gemini Notebook routes each query to the right model tier, up to 40% lower latency on everyday questions
→ Stitch Rapid builds playable wireframes of a full product flow in seconds, the clip below is 1x speed
→ Search uses it as a safety check on MCP tools

If your agent pays a chat model to answer yes or no at every step, this is the cheap router to swap in first

[Styxar @styxar]

@KaiffYT One studio was honest about an AI bug fix and the comments turned into a trial. The next studio will just call it a happy accident