meander_water

> the model has its own emergent guardrails that sometimes cause it to push back on legitimate security research requests. But as we found, these organic refusals aren’t consistent - the same task, framed differently or presented in a different context, could produce completely different outcomes as illustrated in the examples below.

This was new. I'm surprised that a model specifically designed for security research and gated to professionals is refusing legitimate requests

meander_water··on Show HN: Needle: We Distilled Gemini Tool Calling into a 26M Model

I'm so excited for this, nice work!

Gemma4 edge models were promised to be great for agentic use, but have been really disappointing in all my tests. They fail at the most basic tool use scenarios.

Have you run any tool-use benchmarks for Needle, or do you plan to? Would be great if you could add results to the repo if so.

meander_water··on Supply chain compromise in mistralai Python package

This appears to be part of the same Mini Shai-Hulud campaign affecting Tanstack Router https://www.securityweek.com/tanstack-mistral-ai-uipath-hit-...

meander_water··on If AI writes your code, why use Python?

Sure, but will they download the right version? And will they be inspecting the right files on disk? There's a whole lot more that can go wrong

meander_water··on If AI writes your code, why use Python?

One underrated advantage of using Python or Typescript is that AI agents can inspect the code of installed dependencies.

This means you don't have to muck around with supplying the right documentation for each version of each dependency, or worry about hallucinated interfaces (at least with the latest models).

In the past you'd have to dig through a foreign codebase manually to figure out why a documented interface for a dependency is not working as expected, but frontier models automate that quite well.

meander_water··on Postmortem: TanStack NPM supply-chain compromise

I don't understand why people were voting this comment down in the issue page

meander_water··on LLMs corrupt your documents when you delegate

> We find that models are not failing due to “death by a thousand cuts” (i.e., many small errors). Instead, they main- tain near-perfect reconstruction in some rounds, and experience critical failures in a few rounds, typically losing 10-30+ points in a single round trip

> We find that weaker models’ degradation originates primarily from content deletion, while frontier models’ degradation is attributable to corruption of content.

I think we largely already knew this. This is why we fudge around with harnesses and temperature etc.

meander_water··on People Hate AI Art

Agree. All of the major AI model labs have designed their user interfaces in entirely the wrong way.

Prompting via text alone is a really bad way to generate images. Ideally you want Canny Control to draw an outline of the image with elements in the exact locations where you want them. It's why comfyui is so great.

The ability to edit images and specify regions in the image for the prompt is a step in the right directions though. ChatGPT and Gemini have this.

meander_water··on People Hate AI Art

I think one of the reasons for sloppy images is that non-artistic people don't have the vocabulary to describe images to be produced in interesting styles.

Yes, you can do image-> text on existing styles, but something always gets lost in translation.

Midjourney probably has the best baseline, and --sref is a really easy way to differentiate

meander_water··on Principles for agent-native CLIs

Agree with JSON. But, surprisingly html and latex perform slightly better than markdown for more complex tables.

Check out this paper - https://arxiv.org/abs/2506.13405

meander_water··on Should I Run Plain Docker Compose in Production in 2026?

Surprised they didn't mention docker compose secrets - https://docs.docker.com/reference/compose-file/secrets/

meander_water··on Stitch together lots of little HTML pages with navigations for interactions

Isn't this just HATEOAS as espoused by libraries like htmx, datastar, hotwire etc.

https://htmx.org/essays/hateoas/

meander_water··on The Prompt API

Thanks for the insider info! Do you know if there are any published benchmarks for Nano 3?

meander_water··on The Prompt API

This looks like it uses Gemini Nano under the hood. But the latest Gemma4 E2B and E4B models appear to be much better, so you'd probably be better off deploying quantized versions through an extension for now.

- Gemini Nano-1: 46% MMLU, 1.8B

- Gemini Nano-2: 56% MMLU, 3.25B

- Gemma4 E2B: 60.0% MMLU, 2.3B

- Gemma4 E4B: 69.4% MMLU, 4.5B

Sources:

- https://huggingface.co/google/gemma-4-E2B-it

- https://android-developers.googleblog.com/2024/10/gemini-nan...

meander_water··on Our newsroom AI policy

I think most labs actively create synthetic data using existing model as part of the mix for the pretraining stage for their next model.

Would love to know exactly what the latest process is to keep slop out of training data.

meander_water··on Technical, cognitive, and intent debt

Unfortunately large parts of the paper that he linked to from the Wharton school is entirely AI generated, and yet to be peer reviewed.

I realize that most researchers use AI to assist with writing, but when the topic of your paper is "cognitive surrender", I struggle to take any content in there seriously.

meander_water··on This year’s insane timeline of hacks

The crazy part is that none of this is unexpected.

This was exactly the reason why GPT-2 was restricted for general release in 2019.

Check out section 4 - https://cdn.openai.com/GPT_2_August_Report.pdf

meander_water··on Stanford report highlights growing disconnect between AI insiders and everyone

Industrial Revolution? We're still here.

meander_water··on Stanford report highlights growing disconnect between AI insiders and everyone

Zeppelins are another notable one

meander_water··on I ran Gemma 4 as a local model in Codex CLI

I would have liked to see quality results between the different quantization methods - Q4_K_M, Q_8_0, Q_6_K rather than tok/s

meander_water··on Ask HN: What Are You Working On? (April 2026)

I'm building this mostly to scratch my own itch

https://findsubstack.com

It's a newsfeed constructed from 130k substack RSS feeds but limited to the past 24h.

Its helping me discover writers other than just what the algorithm gives me.

meander_water··on Ask HN: What are you working on? (April 2026) (Non AI)

I'm building this mostly to scratch my own itch https://findsubstack.com

A newsfeed for Substack posts from the past 24h. Its helping me discover writers other than just what the algorithm gives me.

meander_water··on Ask HN: What are you building that's not AI related?

https://findsubstack.com

I'm building this mostly to scratch my own itch.

A newsfeed for Substack posts from the past 24h. Its helping me discover writers other than just what the algorithm gives me.

meander_water··on Ask HN: What are you building that's not AI related?

Whoops didn't know about this, should have searched first

meander_water··on Project Glasswing: Securing critical software for the AI era

I think you misunderstood, I do think it's real. I just think they're being disingenuous that this is a new threat. This is the same company that reported that their models were being used by a state actor to perform exploits in real-time - https://www.anthropic.com/news/disrupting-AI-espionage

They know how to run a good marketing campaign.

meander_water··on Taste in the age of AI and LLMs

Not only is the article AI generated, it's recycling a shallow pov that hundreds of other people are just copying.

Just google "taste is the new moat"

Doesn't deserve to be on the front page.

← PreviousPage 2 of 7Next →