Tired of coding agents finding creative ways to cheat tests?
I just shipped an update to Disciplined Vibe Coding featuring a blind-reviewer mechanism (inspired by Google's Elephant/Goldfish model) and hardened hooks to keep the agent honest.
Works with Hermes or standalone harnesses. Try it out, break it, and send PRs: github.com/bmtrnavsky/dis…
I just shipped an update to Disciplined Vibe Coding featuring a blind-reviewer mechanism (inspired by Google's Elephant/Goldfish model) and hardened hooks to keep the agent honest.
Works with Hermes or standalone harnesses. Try it out, break it, and send PRs: github.com/bmtrnavsky/dis…
14
0
8
189
Brad Trnavsky retweeted
Weekend hack: HackerTube. I was annoyed with YouTube's recommendations so I had Claude filter down 43M comments over 20 years of HN to 6,454 videos that keep being recommended. *Significantly* better viewing experience, ~2 hours with Opus 5.5:
hackertube.org
- Excellent (state is kept in local browser session)
- Thumbnail is a representative frame from middle of the video instead of the super duper A/B test hyper-optimized user uploads
- Underlying data from the comment analysis as one zip file, point your Claude at it and ask for your own version
- HackerTube Longs ✊
A bit more on the process:
I use Claude Code (desktop app for Mac), Opus 5.5 on Medium effort as my default. I only talk to Claude and have it do everything, the only "manual work" I helped with in tandem was in the browser (reminiscent of MenuGen): 1) I had to create some kind of API key from Google Cloud UI to query YouTube for video metadata, 2) I booted up a VPS box for serving and 3) I bought the domain name. Once the box was up I gave Claude the IP and told it update, secure, harden, deploy and optimize the code, and it did the whole thing.
Many awesome projects that would have normally taken weeks/months even can now be few-shotted for breakfast. What else are we not thinking of.
hackertube.org
- Excellent (state is kept in local browser session)
- Thumbnail is a representative frame from middle of the video instead of the super duper A/B test hyper-optimized user uploads
- Underlying data from the comment analysis as one zip file, point your Claude at it and ask for your own version
- HackerTube Longs ✊
A bit more on the process:
I use Claude Code (desktop app for Mac), Opus 5.5 on Medium effort as my default. I only talk to Claude and have it do everything, the only "manual work" I helped with in tandem was in the browser (reminiscent of MenuGen): 1) I had to create some kind of API key from Google Cloud UI to query YouTube for video metadata, 2) I booted up a VPS box for serving and 3) I bought the domain name. Once the box was up I gave Claude the IP and told it update, secure, harden, deploy and optimize the code, and it did the whole thing.
Many awesome projects that would have normally taken weeks/months even can now be few-shotted for breakfast. What else are we not thinking of.
1,177
2,448
28,787
2,104,949
Brad Trnavsky retweeted
your Hermes desktop app will soon be able to jump between every gateway you own... without a switch prompt or a blanked session list
@Teknium just merged it! each gateway is dialed at launch, kept open, and redialed if it drops
he expects it in the next full release later this week
@Teknium just merged it! each gateway is dialed at launch, kept open, and redialed if it drops
he expects it in the next full release later this week
Been focusing on the Desktop app this last week or two.
This PR should be welcome news to a lot of you power users.
On next full release later this week I think you'll start noticing the difference in the desktop app getting noticeably nicer.
Still a long way to go, but f UX and design work coming shortly.
4
1
20
1,090
Brad Trnavsky retweeted
Anthropic CEO Dario Amodei just said on live TV what nobody in the tech industry wants to say out loud.
The man building Claude went on the record to expose what Fortune 500 CEOs are privately admitting behind closed doors.
Within the next 1 to 5 years, up to 50% of all nance, law, consulting, and tech are going to vanish.
Two years ago, AI worked like a smart high schooler, and today it operates like a top college graduate that never sleeps, never asks for a raise, and processes documents in seconds.
Anthropic’s own economic data shows 75% of businesses are already using AI to fully automate and replace entire workflows instead of just helping junior employees write faster.
In this 5-minute clip, Dario breaks down the exact shift that is already rewiring the corporate ladder
Basic Prompting → Autonomous Workflows → AI Operators → 1 Person Replacing A 15-Person Department
Using AI to fix your grammar or summarize a PDF will not save your paycheck anymore.
Companies are about to fire the people who do routine digital work and pay 10x more to the people who can wire AI agents straight into their business operations.
90% of white-collar workers will sit around watching AI take their industry, while the 10% who learn to orchestrate AI right now will write their own paychecks for the next decade.
This 5-minute video packs more career-saving truth than any $4,000 AI career course.
Bookmark this and watch it today before your boss does.
The man building Claude went on the record to expose what Fortune 500 CEOs are privately admitting behind closed doors.
Within the next 1 to 5 years, up to 50% of all nance, law, consulting, and tech are going to vanish.
Two years ago, AI worked like a smart high schooler, and today it operates like a top college graduate that never sleeps, never asks for a raise, and processes documents in seconds.
Anthropic’s own economic data shows 75% of businesses are already using AI to fully automate and replace entire workflows instead of just helping junior employees write faster.
In this 5-minute clip, Dario breaks down the exact shift that is already rewiring the corporate ladder
Basic Prompting → Autonomous Workflows → AI Operators → 1 Person Replacing A 15-Person Department
Using AI to fix your grammar or summarize a PDF will not save your paycheck anymore.
Companies are about to fire the people who do routine digital work and pay 10x more to the people who can wire AI agents straight into their business operations.
90% of white-collar workers will sit around watching AI take their industry, while the 10% who learn to orchestrate AI right now will write their own paychecks for the next decade.
This 5-minute video packs more career-saving truth than any $4,000 AI career course.
Bookmark this and watch it today before your boss does.
64
121
562
219,229
Brad Trnavsky retweeted
Customer research gets interesting when someone says, "We already have a spreadsheet for that."
Your SaaS competitor may have no logo, no sales team, and 47 tabs.
Your SaaS competitor may have no logo, no sales team, and 47 tabs.
4
1
1
167
Brad Trnavsky retweeted
"Assessing AI Consciousness" is now available as a book: hutter1.net/ai/bayesconsc.… It develops a Bayesian framework separating theoretical credence from empirical evidence, providing a precise, scalable method for calibrated assessments (with online tool and code): ai-cognition.org/cacophony-tool/ Happy Halloween.
10
83
349
15,955
Brad Trnavsky retweeted
Quantum computer designers just discovered that the “bugs” in their chips can be turned into the feature that makes them scale.
Chalmers physicists showed that microwave crosstalk, the unwanted whispering between neighboring superconducting qubits when you try to talk to only t fixed by the usual trick of tuning a coupler until the qubits stop exchanging energy. Instead, it vanishes through pure destructive interference between several tiny capacitive sneak paths that designers normally treat as fabrication imperfections.
On their test chip the quietest point sat roughly 300 MHz away from the textbook “zero coupling” frequency. The actual cancellation comes from attofarad-scale parasitic capacitors (a few dozen of them, each smaller than a single atom’s diameter in effective size) arranged so the drive signal arrives at the victim qubit from multiple routes with exactly opposing phases. The waves cancel. The unwanted excitation disappears.
The practical surprise is that floating transmons make this alignment far easier to manufacture than grounded ones, the null points shift in the same direction, giving lithography real tolerance instead of demanding impossible precision. The same geometry can also line the classical crosstalk null up with the quantum ZZ null, so two different error sources get suppressed at one operating point.
It’s the quantum equivalent of discovering that the noise canceling headphones work better when you deliberately leave a few carefully sized holes in the earcups. The “parasites” were never the enemy. They were the missing knobs.
Chalmers physicists showed that microwave crosstalk, the unwanted whispering between neighboring superconducting qubits when you try to talk to only t fixed by the usual trick of tuning a coupler until the qubits stop exchanging energy. Instead, it vanishes through pure destructive interference between several tiny capacitive sneak paths that designers normally treat as fabrication imperfections.
On their test chip the quietest point sat roughly 300 MHz away from the textbook “zero coupling” frequency. The actual cancellation comes from attofarad-scale parasitic capacitors (a few dozen of them, each smaller than a single atom’s diameter in effective size) arranged so the drive signal arrives at the victim qubit from multiple routes with exactly opposing phases. The waves cancel. The unwanted excitation disappears.
The practical surprise is that floating transmons make this alignment far easier to manufacture than grounded ones, the null points shift in the same direction, giving lithography real tolerance instead of demanding impossible precision. The same geometry can also line the classical crosstalk null up with the quantum ZZ null, so two different error sources get suppressed at one operating point.
It’s the quantum equivalent of discovering that the noise canceling headphones work better when you deliberately leave a few carefully sized holes in the earcups. The “parasites” were never the enemy. They were the missing knobs.
2
1
2
80
Brad Trnavsky retweeted
THIS THING IS CLEAN AF! 🪽
The Hermes Mobile app is in beta right now.
If you weren’t at Nous Con, here is your very first look!
Well worth the wait! 🔥
The Hermes Mobile app is in beta right now.
If you weren’t at Nous Con, here is your very first look!
Well worth the wait! 🔥
🚨 I MISSED NOUS CON. BUT I DIDN’T MISS THIS!
While everyone else is living it up at Nous Con, I somehow got my grubby little paws on the Hermes Agent MOBILE beta. 🪽📱
HERMES. ON MY IPHONE.
That Install button is about to have a VERY bad night. 🤪
70
20
489
38,123
Brad Trnavsky retweeted
Getting access to the Hermes Agent Mobile app beta while being at NousCon 2026 was shocking!
First impressions of the mobile app:
- sleek design
- feels smooth
best of all its has room for bunch of cool new features to be added in the future!
I am so excited to be part of 🪽@NousResearch
First impressions of the mobile app:
- sleek design
- feels smooth
best of all its has room for bunch of cool new features to be added in the future!
I am so excited to be part of 🪽@NousResearch
40
9
230
21,376
ere is a Hermes plugin / coding agent I built. It’s my daily driver for coding. It’s inspired by Superpowers, G Stack, Ponytail, and a few others. I took the best parts built a plugin to enforce the workflow and wrote all the skills. I think it’s pretty cool. It’s a Hermes plugin but I have instructions for using it with other agents. Take a look and let me know what you think. Take a test drive and if you break it send a PR and I’ll get it fixed. github.com/bmtrnavsky/dis…
0
0
6
757
Here is a Hermes plugin / coding agent I built. It’s my daily driver for coding. It’s inspired by Superpowers, G Stack, Ponytail, and a few others. I took the best parts built a plugin to enforce the workflow and wrote all the skills. I think it’s pretty cool. It’s a Hermes plugin but I have instructions for using it with other agents. Take a look and let me know what you think. Take a test drive and if you break it send a PR and I’ll get it fixed. github.com/bmtrnavsky/dis…
0
0
1
60
298 questions, 2,000 arXiv papers, one result: letting the agent search again gained 1.2 points, pure noise. Agentic RAG did not retrieve better evidence than a single careful pass. More rounds did not help; the first pass did the work. Are you adding retrieval rounds or fixing the first one? towardsdatascience.com/does-agentic-r… #RAG #AI
7
0
4
154
49.6% to 96.5% on a 90-question warehouse eval. Switching the analytics agent from Sonnet to Opus moved accuracy to 88.2%, then clarifying business metrics and killing ambiguous terms added the rest. The model swap helped, but the glossary did nearly as much. Is your agent failing on model quality or on definitions nobody wrote down? benchouse.ai/blog/getting-a… #Analytics #AI #operations
1
0
1
22
I've been working on my RAG today... I discovered some issues and had to make changes, so while I was in there, I re-engineered the whole thing lol. Please take a look and let me know what you think: github.com/bmtrnavsky/sci…
0
1
1
60
Sonnet 5.5 at max reasoning burns 193,000 tokens per task, 60 percent more than Opus, and costs more per task. Max effort as the default looks expensive from here. Do you measure cost per task on your agent runs? anthropic.com/claude-haiku-5… #Claude #AICosts
3
0
0
61
Why are people fascinated with these little virtual offices with AI avatars walking around and working at desks? What does it do for you? Is it more fun? Just cool? Or is it as lame as I think it is?
0
0
0
71
This is waiting for approval in the Hermes Plugin Repository. I built it because most of what was out there was either far too complex for small projects or too simple and didn't require discipline, so I built what I needed. You don't need a computer science degree to ship solid software, but you do need engineering discipline.
I packaged my personal Hermes agent setup into disciplined-vibe-coding. It uses custom persona rules, multi-stage skill gates, and shell hooks to keep both you and the LLM honest from intake to verification.
Available now on GitHub: github.com/bmtrnavsky/dis…
I packaged my personal Hermes agent setup into disciplined-vibe-coding. It uses custom persona rules, multi-stage skill gates, and shell hooks to keep both you and the LLM honest from intake to verification.
Available now on GitHub: github.com/bmtrnavsky/dis…
3
0
4
134
Stanford’s STORM asks how to write an article. Business leadership asks how to make the call.
I built STORM-Fusion-Research-Pipeline, an adaptation of Stanford’s perspective-panel research that replaces passive synthesis with a human-gated decision panel.
Take a tough operational call. Seat the hourly worker, management, finance, marketing, HR, compliance, and the customer. Same problem, seven distinct perspectives. Every seat can run its own multi-model fanout (Hermes MoA or OpenRouter Fusion), a judge highlights the trade-offs, and you make the call. It's live now in the Hermes plugin repository. github.com/bmtrnavsky/sto…
I built STORM-Fusion-Research-Pipeline, an adaptation of Stanford’s perspective-panel research that replaces passive synthesis with a human-gated decision panel.
Take a tough operational call. Seat the hourly worker, management, finance, marketing, HR, compliance, and the customer. Same problem, seven distinct perspectives. Every seat can run its own multi-model fanout (Hermes MoA or OpenRouter Fusion), a judge highlights the trade-offs, and you make the call. It's live now in the Hermes plugin repository. github.com/bmtrnavsky/sto…
0
0
1
35
I've been building so I guess I'll share a couple of my projects.
I hate trying to remember what time it is in UTC, so I built a tool that helps my ADHD brain manage my token spend and avoid accidentally burning full-price tokens on DeepSeek.
I built DeepSeek Whale, a lightweight titlebar widget for Hermes Desktop that tracks peak vs. off-peak pricing in real time.
🔵 Blue whale = 50% discount ⚪ Grey whale = Peak hours
Zero API keys, zero network calls, pure local clock math.
Live now in the Hermes plugin catalog: deepseek-whale
github.com/bmtrnavsky/dee…
I hate trying to remember what time it is in UTC, so I built a tool that helps my ADHD brain manage my token spend and avoid accidentally burning full-price tokens on DeepSeek.
I built DeepSeek Whale, a lightweight titlebar widget for Hermes Desktop that tracks peak vs. off-peak pricing in real time.
🔵 Blue whale = 50% discount ⚪ Grey whale = Peak hours
Zero API keys, zero network calls, pure local clock math.
Live now in the Hermes plugin catalog: deepseek-whale
github.com/bmtrnavsky/dee…
1
0
1
126