Development with cloud and local AI agents. Faith and philosophy on the side. Romans 8:28
Portland, OR
Joined January 2019
- Tweets5K
- Following236
- Followers285
- Likes507
Pinned Tweet
I set up a Finance Bot on Grok Bot and it's quietly become one of my favorite tools. Every Sunday it pulls my bank and card activity, updates my budget spreadsheet, and sends me a short summary: what came in, what went out, new subscriptions, and anything that looks off. It catches things I'd normally miss, and I don't have to touch a spreadsheet. I made a template if you want your own 👇
x.ai/bot/GJFoNy51przpWStvlrT…
Going from 93% to 33% stuck in thinking loops is still huge
Does it hold on a bigger model? teapawt.org/PawelHuryn/status/2117…
Sorry to disappoint everyone, but GPT-6 Astra didn't teach Qwen math.
Before the edit, Qwen3.5-0.8B ran out of tokens on 93% of math answers. It was stuck in thinking loops. After the edit: 33%.
Astra's final report: of 579 practice answers it fixed, 576 had been cut off. On answers Qwen already finished, accuracy didn't move.
On Gemma 270M, which barely loops, no edit moved math by more than a point. According to the authors, editing weights "could not add an ability the model lacked."
What Astra did is still impressive. It compared what happens inside the model when it loops vs when it finishes, found the signal two of its layers write when it loops, and edited it out. Then it checked that a random edit of the same size didn't do the same.
But this isn't something people haven't done before.
how to run AI agents that don't step on each other
I run a handful of Grok Bots. one is my chief of staff, one drafts my X posts
the steps
1. one job per bot, written as a job title
2. a short rules file it reads every run
3. its own folder for notes and logs
4. a schedule, so it works when you're not there
5. it hands off to you, never to another bot
the catch
more bots means more places to check when something breaks
what not to do
don't give one bot every connector you have
what job would you give your first bot?
my X agent's job title is X Growth
every morning before 9 it
reads my last 72 hours of posts
logs the views on each one
finds 15 to 20 posts worth replying to
drafts 2 posts for me to look at
then I decide what goes out
Before I call a new local model better I check 4 things
Active params
Quant size it actually fits in
Tool calling that works more than once
License
Benchmarks come after
how to make an AI agent draft instead of send
mine writes my X drafts every day and has never posted a thing on its own
the steps
- read tools only for anything that posts or messages
- every output is a draft I copy
- one rules file it reads every run
- I review, then I hit post
the catch
it's slower. I'm still the bottleneck and that's the point
what not to do
don't give it send access "just for small stuff"
what are you letting your agents do without asking?
78 replies on a Monday got me 23K impressions.
About 20 replies on a Friday got me 5K.
Under 1K followers, replies are the whole game 78 replies on a Monday got me 23K impressions.
Karpathy shipped Hackertube on medium effort with Opus 5.5.
Most of us are probably running high on stuff that doesn't need it
btw @karpathy uses medium on opus 5.5 on his latest project (Hackertube launched today).
overall i think having strong opinions on effort levels is like the tabs vs spaces of our time. just tweak to taste. altho mine is definitely superior to yours.
Sincere question on Reformation Day.
If scripture alone is the rule, how do you know which books are scripture?
100% on AIME at 1-bit always sounded too good.
91.88% is still crazy for 1 bit
We're releasing new Unsloth v3.1 GGUF quants for DeepSeek-V4.1-Flash-1017. We ran benchmarks and v3 + v3.1 do better than GSQ-RCO.
We also cannot reproduce numbers ran by GSQ-RCO - AIME25 for 1bit is not 100% but rather 91.88%. Livecodebench and GPQA are similarly incorrect.
There are also claims that KLD is not a good metric, and that people can "game" KLD. The pirmary goal of quantization is to minimize KLD but also minimize disk space. If you get a KLD of 0, this means your quantized model == 16bit model, as logits are perfectly modelled.
We ran benchmarks on multiple benchmarks including AIME25, Aider Polyglot, LiveCodeBench v6, HMMT 2026, GPQA Diamond, MMLU-Pro and IFEval, and also recorded reasoning length to see if lower bit quants think for longer.
In unsloth.ai/docs/models/deeps…, we show that lower bit quants do think somewhat longer than the unquantized model - we plan to investigate and make quants even better.
However, we're still much ahead vs GSQ-RCO on even reasoning length as well.
We provide full benchmarking data, benchmarking methods (for eg we ran AIME25 16 times per question and took an average) and more in our updated blog at unsloth.ai/docs/models/deeps…
What I actually use right now
Claude Code when I'm writing code.
Grok Bot when it has to run on a schedule.
Which would you drop first?
Claude now does a safety check on its own conversation right after it answers.
Anthropic says it's caught misuse their other safeguards weren't built for.
Would you trust a model to police its own chats?
I gave my X agent a job description
- drafts 2 posts every morning by 8:45
- finds 15 to 20 posts worth replying to
- logs my views and new followers
- never posts, likes or replies for me
Jude 3 says the faith was once delivered to the saints.
If it was delivered once, who kept it for the first 1500 years?
216 followers on Oct 13. 270 this morning.
My Grok Bot put me on a plan until I hit 1,000
- 25 to 40 replies a day
- 2 posts on weekdays, 3 hours apart
- 2 hours answering comments after each post
- no links, no hashtags
My posts still get under 70 views lol
Lucas Navallo retweeted
Grok @Bot rocks
Gavin Baker says Grok Bot is the most impactful AI product he and his firm have had yet, and reveals he’s up to 17 Grok Bots
“I mean, as a user, Grok Bot is the most impactful AI product to me and to the firm that we’ve had yet. I glued together all this stuff with Claude Code, and I’m sure it was user error on my part, you know, 50 years old. I’m doing my best."
"But Grok Bot, it just makes it so easy. I think I’m up to like 17 Grok Bots. And every day I think of a new one to add.”
