HN user

tacotime

243 karma
Posts0
Comments123
View on HN
No posts found.

I doubt it is possible to mathematically prove much inside of a black box of billions of interconnected weights. But at least in the narrow case of the strawberry problem, it seems likely that LLM inference could reliably recognizing that sort of problem as the type that would benefit from a letter counting tool call as part of the response.

I wonder if using the character itself in the directions, instead of the name for the character, might help with this.

Something like, "Forbidden character list: [—, –]" or "Do NOT use the characters '—' or '–' in any of your output"

I could imagine a vision-enabled transformer model being useful to create a customizable “reading mode”, that adjusts page layout based on things like user prefs, monitor/window size, ad recognition, visual detail of images, information density of the text, etc.

Maybe in an alternate universe where every user-agent enabled browser had this type of thing enabled by default, most companies would skip site design all together and just publish raw ad copy, info, and images.

Hacker Typer 2 years ago

Triple shift pops up a red "Access Denied" message in the middle of the screen. Alt just shifts focus to the browser's toolbar, at least in Chrome.

I have been in a car accident that, in the instant of impact, felt indistinguishable from an explosion, too fast to process anything. I imagine the massive pressure change would cause near-instant incapacitation if not death.

HDR QR Code 3 years ago

On my macbook with vscode on the left and chrome on the right. When I click on this link with the right side of my screen covered up, I can visibly see vscode on the left get a couple of shades darker (animated over a second or so). So it looks like it's not only a question of hidden max brightness, it looks like the device adjusts the whole screen to enhance contrast for the HDR content as part of the strategy.

Strictly work related ones probably still would, but maybe at a reduced rate.

What if there was the promise of a friendly AI used to filter out all non-work related discussions?

Then a chatGPT model fine tuned based on previous team discussions, that automatically replies in the team chat any time it has some especially high confidence that it's generated a good answer to a new question.

On my team, there is a lot of "tribal knowledge" that is known to part of the team, and buried in Microsoft Team's chat history somewhere, but having to use Team's search for anything is always the last resort. Maybe add a process where you have to wait for a team member to "like" the AI generated post before the person who posed the question uses the AI answer.

A raccoon can definitely kill and eat (part of) a chicken so I’d imagine a crow could be on the menu. Or a coyote. Or a possum or bear. There are vids of deer eating birds and other animals.

At least on Windows, you should be able to modify the target property of the program’s shortcut to add any desired flags. Then use that shortcut to launch, or drop it in the startup folder. On Linux idk, maybe an alias?

Someone in my family, who is now in a memory care facility, has dementia. We tried to avoid it as long as possible, but when we got a call from the police saying that she was found wandering around barefoot in the snow next to a busy road in the middle of the night while hallucinating about “people coming to get her”, it felt like it was time to reconsider. She’s on meds that control the hallucinations now, but it would still be unsafe for her to leave unsupervised, no matter how much our culture changes. How do you help a stranger who doesn’t know where or who they are, and can’t hear too well to start with? She sure does still remember me and can dance to the Beatles, but I would not be comfortable trusting her welfare to a group of even the most compassionate strangers.

Not in this case. Visitors to this site are only pulling data from a torrent, not hosting any data for other visitors. If you want to help host the torrent then you'd be seeding the whole database by default. Torrents are split into blocks and there is no easy way to choose to only download/seed specific blocks of a single file.

Maybe something like Chia (chia.net) that "farms" from pre-computed "plots" as proof of work? There is still an energy intensive cost to the initial plot creation phase, but it is soon amortized by the low cost farming phase, where the precomputed proofs begin to do their work. As SSD storage is expected to become less $/TB than HDD storage in about 5 years, this could make for a very energy efficient PoW scheme.

I don't understand enough to know if it's a design decision or a technological limitation but the barrier to entry for something like chia also seems much lower, hopefully democratizing the ability to participate in the creation/validation side of the network. Chia requires a single ~100GB plot file (and 332 GiB scratch space when creating the plot) to start farming, which is a pretty low bar. I don't have to buy an expensive special purpose bitcoin ASIC and I don't have to stake the equivalent of 68,416.96 United States Dollar (32 eth) to get started and start contributing.

I’m a regular laptop user and I often use this trick. In Windows you can ctrl+click the back button to pop open a tab for the previous url, I’d guess command+clicking the back button does the same on a Mac.

Is there a way to get this working using the Microsoft Store version of Windows Terminal? I've tried editing %appdata%\Local\Packages\Microsoft.WindowsTerminal_8wekyb3d8bbwe\LocalState\settings.json and nothing changed.

I tried to update C:\Program Files\WindowsApps\Microsoft.WindowsTerminal_1.5.10271.0_x64__8wekyb3d8bbwe\defaults.json but it requires SYSTEM privileges to edit... I wanted to check this out badly but I am at a loss...