neurostimulant

I thought it's just a marketing gimmick, just like how cars should be safely driving by itself by now.

Where did you test the prompt? GLM 5.3 flash can answer those questions just fine on openrouter: https://i.imgur.com/NbEDzZX.png

Even then, some phones that advertise bypass charging are actually using a software-based implementation, and they can still end up with swollen batteries after being plugged in for extended periods.

When OpenAI was actually still a proponent of open AI...

> Anthropic said in a February posting that it had identified a campaign by Chinese AI startup DeepSeek ...

> It said DeepSeek's operation involved over 150,000 exchanges

That volume seems more like the number of requests 15 employees using Claude Code would generate in a month. It seems too small for a large scale model distillation campaign.

Very interesting. But why has nobody tried to do prompt injection attacks on this AI agent?

I wonder if you could ask Opus 4.7 to port LineageOS for your phone. The LLM must’ve had plenty of training data from the gazillion LineageOS ports for other devices, so perhaps it might actually work (or brick your device).

> Xiaomi has been ironically the pioneer in this field, but their phones are inaccessible in the USA assuming you’re USA based

Xiaomi has virtually stopped allowing users to unlock their bootloaders. They only allow 2000 device unlocks per day, which resets every day at 00:00 Beijing time. You can set an alarm and press the unlock button exactly at 00:00 Beijing time, only to get frozen out due to the thundering herd effect and fail to secure an unlock quota. Then, when you try again at 00:01, you just get the dreaded "quota limit reached, try again tomorrow" message.

Isn't high memory pressure will result in high memory paging which will wear down the internal ssd?

Maybe we should go back to kitchen-sink frameworks so most functionality you need is covered by the fat framework. I'm still using django and it keeps my python project's dependency relatively low :)

I'm able to sweet talk the gemma-4-e2b-it model in an iphone 15 to solve a hcaptcha screenshot. This small model is surprisingly very capable!

> I wish this mission took greater risks

It's already risky enough: https://idlewords.com/2026/03/artemis_ii_is_not_safe_to_fly....

Well, at least that confirms they weren't lying when they said all recent updates to claude code were made by claude. You certainly won't do this stuff if you were writing the code yourself.

Just check the leaked code yourself. Two biggest areas seem to be the `utils` module, which is a kitchen sink that covers a lot of functionality from sandboxing, git support, sessions, etc, and `components` module, which contains the react ui. You could certainly build a cli agent with much smaller codebase, with leaner ui code without react, but probably not with this truckload of functionality.

I think that only happens when you use an old browser without native sha256 support (so it fallback to pure js implementation of sha256).

> The IPs of these bots here actually do not come from datacenters or VPNs most of the time; the overwhelming majority come from residential and mobile networks. Asian and Indonesian countries are where nearly all of them reside.

It's really awful as an indonesian, my indonesian isp regularly got blocked by HN as well :(

Maybe they shouldn't get rid of subpixel antialiasing. At least this problem wouldn't be as painful if we still have that feature.

But isn't it normal for people who work on AI stuff to use LLMs for everything? They are very enthusiastic about AI so naturally they'll use it on everything they can.

The processor is still fast enough, the ram is too small for modern web apps.

Skills are only loaded when you need them, so you’ll probably use fewer tokens overall compared to MCP servers or including them manually in your main AGENTS.md/CLAUDE.md file, which are always loaded in the system prompt.

> Earbuds don't have displays, so the only way to transfer data to/from them is either via bluetooth, or the UART contact pads. Bluetooth is pretty slow, you'd be lucky to get a consistent 1mbps connection, UART is easily the better option.

Does this means you can run a doom instance on each bud? Is it viable to make a distributed app to use the computing power of both buds at once?

How long until antivirus vendors start flagging zig binaries as malware (like they did to golang binaries a while ago)?

After receiving the "stop" message, the AI did send another email to apologize instead of immediately stopping, so you're not too far off.

Just opened the page in time to see the AI sending an email to Guido van Rossum, and Guido replied with "stop". Wild.

My og chromecast frequently crash (I suspect due to running out of ram when casting contents) so I ended up replacing it with the chromecast ultra. The og chromecast is still usable if you can tolerate the crashing, but I imagine it's going to get worse as developers keep updating their cast players (basically a webpage that plays video) with new features that consume more and more ram.

Forgot to add immich_postgres which uses ~222 MB of ram. If you already has a postgres running somewhere, you might be able to reuse it as long as its version >= 14 with pgvector installed.

I think you can disable machine learning stuff in the settings. And yes, you can control when the sync happen, or even have it only triggered manually and never automatically.

My instance when not running any photo/video processing job:

- immich server: cpu: 0.51% ram: 753MiB

- immich redis: cpu: 0.48 ram: 19.64MiB

- immich machine learning: cpu: 0% ram: 209.3MiB

It has modern features that requires relatively heavy processing such as facial recognition, finding similar images, ocr, transcoding videos, etc. I think it only needs those computing resources when you upload new images/videos.

Vibe code fixer seems to be a viable job soon

Zed is actually pretty good at this.