bobkb

Happy to see an open weight model ! This has all the right ingredients for success.

Would love to collaborate

Thank you! Jetstream looks useful. For a platform which is production our Redis costs are getting prohibitively high. I was on the look for alternatives and perhaps in the software reliability tool [0] I am buildin try out Jetstream.

[0] https://github.com/bobinson/vulture

Wow!

I have a stupid question - as I understand NATS works very well as a “message” pipe/bus. Anyway to get Redis type cache functionality as well ? Is it something possible ?

We are in the same decision making stage and trying to choose between rabitMQ, Kafka, NATS and some solutions built on top of Redis.

Did NATS eventually worked well?

Not sure whether this is the right step. Rather steps must be taken to ensure reliability of the software irrespective of whether it’s “written” or “generated”. The code generation based on “rules” is not new after all.

The leader boards based on token usage happened in our org for a month. Then we managed to convince the board that what matters is the reliable software shipped. Now we are back to DORA metrics.

How to prove this is indeed the system prompt against a certain timestamp (in the past ) ?

Very useful. The whisper setup is something similar to what we have been using. The LLM setup though is outstanding.

This gave me shivers

Will it be accessible to anyone ?

This type of attack is going on for few years now. I had 2 in my credit.

Some details https://freebird.in/malicious-code-source-code-shared-via-jo...

Will cursor launch a CLI tool like Claude/codex/opencode/pi ?

I assume your idea is, if the spec and the proof is verified the code generated is good enough as well ?

I have been testing formal verification methods with multiple products. It will be great to also understand more about what’s tried and how it was done. For example attempting to verify the spec is what I have been trying to implement.

In an interesting coincidence I ended up watching Person of Interest S4 E5 while reading the announcement. The series showed some code supposedly belonging to to an AI.

Fable 5 said the first screen shot is from “ IDA Pro’s Hex-Rays decompiler” and a windows driver. The second screenshot triggered the safety guard rails and pushed me into Haiku.

Apparently the code is Windows driver code.

It’s impossible to write a spec that’s not ambiguous , complete and correct in natural languages. Thus prompts will always generate unreliable software.

IMHO even if we are using auditing tools I believe we must use deterministic tools for critical analysis like this. Such rule and pattern based systems may not scale beyond certain point but they can be accurate.

At work we are now in the process of migrating away from Figma. We had spend years perfecting our Figma based design workflow. Currently we are moving all the designs into the code itself using Storybook. The gap currently is reviews and feedback which is addressed by Chromatic now.

I tried building a deliberately vague project around managing MCP servers [0]. The purpose was to find what LLMs and agents can do. While the project didn’t reach anywhere I was amazed by how it’s possible to navigate even with no clear direction. The ability of the “glorified auto-complete” system to pull off something this sort was an eye opener for me.

0. https://github.com/bobinson/aop1

False positives from the deterministic audits a very difficult problem to address. Comparing and deduplicating across different methods or LLM audits seems to the only way.

I think these audit tools can look beyond just security and can look for compliance audits as well. The ability to audit real targets in staging environments makes it easy to identify issues.

Very interesting.

I have working on and using a similar tool for a while now :

https://github.com/bobinson/vulture

I have been struggling with false positives and using Claude + MCP as a poor man’s audit tool. As of last few days found better result with nvidia hosted models.

When will npm issues stop ? This has become a big pain !

That’s impressive!

On the sheer performance it’s comparable to Opus ?

Agreed the marketing angle. But beyond the marketing angle what seems to matter is the access to data - look at Seedance , various Kling models etc which are far ahead of others.

The trouble is the lack of training available to these models compared to the ones like Seedance and Kling who seems to be tapping into their unlimited video inventory. Many models like LTX is technically good but when it comes to slightly different camera movements or the subject interacting with objects they struggle. For a recent example we had to use sample videos generated by closed source models and then use the same for final video.

Nostalgic alumni here :)

This is a manual rewrite or auto generation ?

While those who are hands on is realising the limits and issues with vibe/context engineering/agentic engeering/buzz-word-of-the-week the businesses and pushing hard on the buzz words. It’s high time we start looking at ways to live with the new reality and figure out ways to ensure software reliability.