Murr

No guru shit. Just experiments & progress.

[Murr @murrtlen]

@FENRO0 The catch with a cheaper breaker model: it only finds the failures it can imagine. Overfit and slippage, sure. Regime shifts it has never seen in data, not really. I'd add one dumb rule on top: forward test on live data with tiny size before anything ships.

[Murr @murrtlen]

A startup just put its $5.5M seed round behind one claim: the next leap in AI coding won’t come from OpenAI or Anthropic.

Their answer is a feedback loop that learns from you. The product is Beacon, open source under MIT, from Asymptote Labs.

Here is what the README describes. It records your agent sessions (prompts, tool calls, commands, file activity) from Claude Code, Cursor, Codex CLI and 20+ other harnesses. It evaluates what worked, extracts knowledge, asks you to approve it, then serves it to the next agent through MCP or Agent Skills.

The part I like: data stays on your machine by default. Cloud forwarding is opt-in, and approval comes before reuse.

The part I can’t judge: whether the skills it extracts are any good. The README doesn’t say how it decides what is useful, and some harnesses only capture part of the events.

So I read the pitch as a hypothesis, not a result. Models keep improving without your help. The open question is whether your own session history adds something a bigger model doesn’t already have.

I haven’t installed it. If I do, the test is simple: one repo, one week, and I count how many times I stop re-explaining the same project.

[Murr @murrtlen]

This is getting out of hand: one Grok Bot routine that checks every 5 minutes wakes up 8,640 times a month.
If only about 6 things change a day, 98% of those runs find nothing. You still pay for each one.
Same bot. Same model. Same prompt. Only the clock changes:
Every 5 min: ~$397 a month
Every hour: ~$33
Every 5 min with a tiny check first: ~$45
Estimates at Grok 4.6 API rates for a run of 20K tokens in and 1K out. Not a measured bill.
The third line is boring and it works on paper. Before the full run, read one hash or one timestamp and stop if nothing changed. You keep the 5-minute reaction time and pay full price only for the ~180 runs that matter.
Two more things from the article worth saving. A night routine that needs your approval doesn’t wait: background approvals expire after about 10 minutes. And only the last 20 runs per routine are kept, so write your own log.
Everyone is busy choosing which bot to hire. I think the first thing to audit is how often each one wakes up.
I re-did the arithmetic from the article below. I haven’t measured my own bill, and “6 changes a day” is an assumption, so check your own run log before copying any of this.

[Sprytix @Sprytixl]

https://t.co/aRj8erlMO9

[Murr @murrtlen]

@coolerwave Worth keeping the word "reported". Stripe confirmed the deal but not the price. NYT says $7.5B, Axios says more than $8B, mostly stock.
The fit makes sense to me though. A model router is a billing meter with models behind it, and metering is what Stripe sells.

[Murr @murrtlen]

Nobody likes a boring guide. Everybody saves it.
A post about a Claude skill that draws a hand-made washing machine you can actually operate got 261K views.
The number that stopped me was not the views.
Bookmarks: about 5K. Likes: about 4.9K.
Saves matched likes. Reposts were 515. Replies were 78.
Do the math on the rounded counts. One save every 52 views. One reply every 3,350 views.
Most AI demos I scroll past get a quick like and nothing else. This one got saved, which is a different action. A like says "nice". A save says "I want to come back to this and use it".
My read, and it is only a read: people do not bookmark a clever prompt, they bookmark something they can touch. Buttons that change a real state. A slider for hand-drawn wobble. A catalog that goes from Fig. 01 to Fig. 12 and keeps growing.
So my next experiment is simple. Same idea, different object, and I track saves against likes on my own post.
One post is one data point, not a rule. If you have a case where saves beat likes, I want to see it.

[Bober_smart @Bober_smart]

https://t.co/WRvaPXgqIa

[Bober_smart @Bober_smart]

https://t.co/WRvaPXgqIa

[Elvar @elvar_official]

https://t.co/NAgH1EYziQ

[Murr @murrtlen]

@EnvolDev The real chart here isn’t quality, it’s time-to-done.A model that’s 2x cheaper but 4x slower isn’t cheaper once an agent is waiting on it

[Murr @murrtlen]

@Aprilzz423 The wobble slider is the smartest detail here. Imperfect lines make people treat it like a toy instead of a diagram. Curious if the skill decides the machine states itself, or you define them per object?

[Murr @murrtlen]

https://t.co/3dSwAt26m8

[Murr @murrtlen]

I tracked one big AI account for 5 days. It opened 8 posts with the exact same line.

Views per post, in posting order, as of today:
788K, 356K, 21K, 78K, 6K, 105K, 10K, 15K.
The top two posts hold 83% of all views. The median post got 49K. The best one got about 128 times the views of the worst.

Same opening line, same account, same week. So the line is not what carries a post. Whatever was inside each one did.

Caveats: these are public view counts I read on Oct 18. The newest posts have had less time to grow, so the tail will move. Eight posts is a small sample, and I can’t see why the algorithm treated each one differently.

What I’m taking from it for my own account: I won’t reuse a hook just because it worked once. I’ll write each opening for the specific post, and I’ll track my views the same way and post the numbers.

[Murr @murrtlen]

Many people are showing what AI built around the flying Roadster. I asked three Claude models to build the car itself, with working thrusters. The real design is not public, so each page is the model’s own guess.

Same brief for all three: a stylized 3D Roadster in the browser, drag to rotate, a THRUST button that lights 10 pink thrusters and lifts the car off the ground. No libraries, one attempt each, no edits.

What came back:
Haiku: 341 lines, 3:05 of agent time. A blocky red coupe with glowing thruster dots.
Sonnet: 310 lines, 2:40. A silver car with sharp pink flames under the body.
Opus: 415 lines, 4:50. A glossy red coupe with a live altitude readout in the corner.
All three ran with zero JavaScript errors in my headless check. Opus also reported its own flaws: a jagged edge where the glass meets the paint, and flames that sometimes draw over the body.
For reference, what is actually published about the real car: 10 thrusters, 600 L of compressed air, a 2028 target. Flight height and flight time are not published, so none of these pages is the real Roadster.
The times are how long each model run took, not a speed benchmark. This is one run per model, so it says little about which model is better at this kind of task.

[Darren @dtzy_88]

DeepSeek V4.1 Flash 1710 的鹈鹕骑车测试 🚲

细节上比 GPT-6.1 Sol xhigh 好多了，但还是干不过 Opus 5.5 max。

就是这骑车的速度实在太慢了，像老大爷慢悠悠蹬三轮，看着莫名好笑 🤣

[Murr @murrtlen]

Gemini 4 Argon built a game around the flying Roadster. I wanted to see the car itself, so I gave a text brief to Claude Opus and asked for a wireframe model. Text only, no photos, one attempt, no edits.

The brief listed 17 required parts: body shell, canopy, four wheels, an air tank and 10 thrusters, at roughly 4.2 x 1.95 x 1.15 m.

What came back: 20 parts, 2,956 line segments, all 17 required parts present, 10 of 10 nozzles and 100% left-right symmetry on my check. Size 4.3 x 2.0 x 1.1 m. It took about 35 seconds.

I ran the same brief on Sonnet and Haiku too. Sonnet made 19 parts in about 32 seconds, Haiku made 18 parts in about 90 seconds. All three hit 17 of 17 required parts and 10 of 10 nozzles. The gap was in detail: Haiku 1,695 segments, Sonnet 2,493, Opus 2,956.

What I can’t claim: the real car’s design is not public beyond 10 thrusters, 600 L of compressed air and a 2028 target. Flight height and flight time are not published. The models never saw the real design, so the nozzle positions are their guess, and the flames in the video are decorative.

Treat it as a spec-following test, not a picture of the real car.

[Ammaar Reshi @ammaar]

I used Gemini 4 Argon to make Subway Surfers for the new flying Tesla Roadster 🚀

Fly over moving traffic, hit the thrusters to barrel roll in the sky, and collect air power-ups to set a high score.

Play + code below!

[Cooler @coolerwave]

88 million people watched this. the video never shows what's under the panels.

so I had Opus 5.5 take it apart from frames of this clip. any part no frame shows got marked unknown.

10 parts traced, 4 unknown, including whatever lifts it.

what should it take apart next?

[Murr @murrtlen]

@HowToPrompt__ 605 deaths to kill one dragon. Impressive coordination, but the real bottleneck for agent teams was never teamwork. It’s the cost of being wrong

[Murr @murrtlen]

4 agents, 1 shared folder, 1 job that matters most: catching the other three.

When the audit agent flags one step, don't restart the whole run. Fix that step and rerun from there.

Most people see one wrong number in a generated lab and throw away the session. That wastes the part that was correct and teaches you nothing about where the error entered.

the loop:
→ every experiment writes a log line: source page, formula, constant used, result
→ the auditor recomputes the result from the paper and writes the gap next to it
→ any gap above your tolerance gets a "flagged" tag and a link to the exact step
→ a repair agent edits only that step, with the paper page and the failing numbers in front of it
→ the auditor reruns the same check on the same input, and the tag clears or stays

stop doing: "regenerate everything" as the answer to one bad number
stop doing: setting no tolerance, "looks close" is not a check

Set the tolerance before the run. Decide what counts as a match while you are still calm, not while you are staring at a result you like.

before: one bad result, a full restart, no idea what changed
after: one flagged step, one local fix, a before and after number for it

[YOUR RESULT HERE: which step got flagged and what the fix was]

prompt in the first reply ↓

[Bober_smart @Bober_smart]

It is becoming f*cking dangerous: AI agents from Claude Code transformed an ordinary academic reading list into an interactive research lab right inside the browser

Boris Cherny, who created Claude Code, gave a dozen agents a stack of 64 research papers with a strict assignment: design a web platform where any user can independently reproduce physical experiments

There you can select a fluid flow model, adjust external conditions, and analyze in real time how the physics of the process changes

The interface allows switching between algorithms right during the scene, inspecting the formulas involved, source scripts, and numerical metrics alongside the graphics

Each link in the team received a clear scope of work:

> rightarrow analysts pull the math, baseline assumptions, and test settings from the texts
> rightarrow coders translate formulas into working algorithms
> rightarrow auditors check the resulting dynamics against the original papers
> rightarrow interface specialists build control panels for every experiment

The danger lies in the fact that a pretty visualization can easily turn out to be mathematically wrong

Therefore, controllers strictly check adherence to mass conservation laws, calculation stability, and accurate matching with the source metrics before launching into production

For every experiment, the interface displays:

the statement being tested
the constants involved
the deviation from the baseline calculation
blind spots requiring additional verification

The resulting sandbox combined 11 approaches and 28 reference scenes. Seven of them matched the original within allowable tolerances, while four remained stuck in "pending revision" status

If you want to run something similar, upload your stack of PDFs into Claude Code and give this command:

Turn these papers into an interactive web complex:

1. Extract working algorithms, mathematical models, and test parameters. Be sure to specify page numbers and highlight gaps
2. Assign different methodologies to isolated agents, unifying them with a shared input field and single measurement tools for correct analytics
3. Set up an independent review stage: let reviewers record bugs, and leave all unverified results clearly marked with warning tags
4. Design convenient controls to tweak parameters, compare methodologies, and run through the source code

It is wiser to start with a single document and a single baseline experiment, gradually scaling up the number of agents

Be sure to pack dependencies and clean input data so that an outside researcher can bring up the entire testbed from scratch with a single button

Full instructions on how to create this can be found in the article below

[Murr @murrtlen]

@thedelost Opus as architect, Haiku and Sonnet as workers. Script-as-plan means orchestration is deterministic and burns no context. Would love to see real numbers on tokens and quality

[Murr @murrtlen]

OpenAI just showed where vibe coding is heading next

the 2026 DevDay keynote is packed with AI agents, Codex, new models, and tools that can take you from an idea to a working product way faster

and the crazy part?

they’re not just talking about the future

they actually show what you can build with these tools right now

if you’re a developer, a vibe coder, or you’re trying to build your first AI product

this is 1 hour you should probably watch