pavelai

LLMs are structurally incapable of abductive "jump" to formulate novel premises required to generate new theories. While they can deduct by using established premises to prove things

Prompt injection allows third-party to inject a system prompt with simple message, by inserting special HTML-like sequence (details below). Some of the issues are well-known and pretty old (almost 2 years for Transformers library)

Issues:

• Ollama: https://github.com/ollama/ollama/issues/15931

• HuggingFace: https://github.com/huggingface/transformers/issues/29279 (labeled as feature request) https://github.com/huggingface/transformers/issues/47822 (with a bug label)

• Gemma4: https://github.com/google-deepmind/gemma/issues/768

The problem is that for tools like Ollama there is no solution except of to fix it by Ollama developers

More context

Why this is important: prompt injections are pretty dangerous and as long as user input is an instruction to the model it could be decided as code injections vulnerability. And in combination with long-memory and multi-agentic runtimes this vulnerability could stay in system for a long time. And it has not been decided as a serious security vulnerability by the global community yet

Temporal Solution

So if you're running local models just make sure to throw an error when there is a special sequence in user input. For Gemma family it is <|turn> and for tiktoken-based models it's <|im_start|>. And would be nice to see more solutions

Disclaimer

I'm not a security expert

The companies were notified 30 days ago about the issue. Only Google responded with a feedback on the issue (swiftly)

This small Klingon speaking language model was trained completely on ESP32. The training took 2 days

Number of parameters: 319K

(Disclaimer) The models is tiny and is not a pocket chatbot. It mistakes and is not capable to support conversation, but that's not the goal of the project

The goals of the project is to bring training to edge devices and it worked out

How can one use it? By using solar panels such device could be turned into autonomous meteorological station

The model was quantized to 8, 4, 2, and 1 bit. Characteristics:

• Q8: 8-bit 1.56 TB, lossless

• Q4: 4-bit, 1.51 TB

• Q2: 2-bit: 861 GB

• Q1: 1-bit, 594 GB

The smallest Q1 model keeps 78.9% accuracy, while being almost 3 times smaller than the original one. Instruction for running the model is in the model's card

"We introduce requential coding, where a teacher model selects training samples drawn from the student's own distribution"

Due to new learning technique the model has achieved better generalization skill without overfitting and memorization. This become possible because of new learning method which made the student model to generate samples for itself. It led to intensive reuse of existing neurons and allowed to encode information in a more dense way

While researchers are calling it a compression, I think it's a retopologization, and Microsoft had tried to do something similar in the past with their Phi model family, which they trained on reduced dictionary and simplified knowledge base first. But it seems like MS' researchers didn't explore this exact way of learning. I believe this should give even better results in the future and this is another small breakthrough moment

Paper: https://arxiv.org/html/2607.11883v1

Repository: https://github.com/shikaiqiu/requential-coding

There are plenty of them on Reddit

Transformers.js allows you to run models right in the browser. The fourth version focuses on performance. The new version has support of WebGPU and it opens new era in browser-run models

Here the demos on HuggingFace: https://huggingface.co/collections/webml-community/transform...

It's just a surprise to see what can be done with the models in browsers today. This demos shows the abilities of the models, and this is the time for creators to bring their ideas and make solutions for real tasks

This release also adds new models to be run in browser Mistral4, Qwen2, DeepSeek-v3 and others. It has limited number of changes, what makes it pretty stable for a major version

While the idea is somewhat new for today's JS. And I see some benefits for this to make single-purpose servers compiled into tiny binaries. I believe it would take some time to make this popular. You should to find the niche where it's required right now. And also I would spend more time working on marketing: explainer, documentation, landing page. For example now the readme looks too ai-written

What's about code and DX: it's not a good practice to export anything using globals, this is what JS world refused to do long ago. It turns your code into a hardly debuggable mess quickly

You forgot to upload a python script or results described in the readme. Without it this is only an idea draft, not a proof of anything

Yep, this is one is a real hero in this list

AtProto is a very unexpected choice to see here. Not because it's not good, but it's just very young.

Why did you chose AtProto?

Obviously it's

* Docker

* WASM

* Rustlang

* Web itself

Hm... Why not to sell it or give to the community, if there is no intension in further work on the project?

I think it's would be faster to build this as a Discord community and then to move to a dedicated app/service. It's too early to fix anything into a profession or term because things are changing too fast. And community would help you to structure the information faster

Omg, it was just right there. Thanks. These projects are incredible examples of developer creativeness. Wanna try it in a browser one day

It would be merged into a valid code. What's about alignment in particular, I'm not sure

MacOS one day?

Builders will be there as long as people do programming. But it's the first step to move to better builders. Or to standard one?

Also it helps to start with typescript faster and easier, and to make the learning curve smoother and maintaining less complex for all developers on all platforms

If one of developers changed function name and the other changes alignment of parameters in the same line, this tool would recognize the changes and merge this line without conflicts. Regular git algorithm would turn it into a conflict because the changes happened on the same line

Ok. We are talking about different things. There are fundamental rules which make technologies to strive or disappear. But there are other fundamental rules for products, features, etc. So you talk about products and I talk about technology.

YCombinator is about products, not technologies, so their example is not suitable here. Agents is a technology. They have fundamental advantage over premade software. They are already here and they are doing valuable work. They will stay for the near future. You can argue against reasons to use particular agents or particular products built around them, or particular features, but not against the whole technology.

The technology is verified, we need to find the way to make it safer and more sustainable. This would require to solve many technical tasks: build safeguards, create eco-friendly energy production, solve explainability, build physical infrastructure, develop new technical standards, etc.

Well, the history of software development shows that everything is changing and don't lasts forever. Golden age of SaaSes now is going to fade (at least for current version and for some time, maybe a decade). Today junior developers can create a fully working SaaS with dozens of features without even knowing how it works. And they would spend a few weeks or evenings to develop it. We are about to see a flood of services and it will make all the business model obsolete. This is what is happening now at Github. You can see a lot of excellent projects which were written by AI, and no one wants to contribute to these projects, because they can do the same project on themselves

You can do a blog about retro programming to find clients which have your stack. I read about group of developers who were writing in Cobol in 2010s. Because this software were in use and still in use somewhere. The guy who wrote about it was a developer with decades of experience, but no one wanted hire him because of the age. Then he found other developers in the same situation and run a company to support software written in outdated languages

You can write old-styled software for those who have nostalgic feelings about good old times, or support solutions written for old platforms. Instead of been thinking about it as of trash, think about it as it is retro