HackerTrans
TopNewTrendsCommentsPastAskShowJobs

oceanparkway

no profile record

Submissions

Career pivot to ML/AI: How did you make the leap?

2 points·by oceanparkway·2 anni fa·0 comments

Unexpected Interface Outcomes per Hour

zachblume.com
2 points·by oceanparkway·2 anni fa·0 comments

Docker Compose inside Docker, autoscaled to zero on Fly.io

zachblume.com
2 points·by oceanparkway·2 anni fa·0 comments

Show HN: Autospec – open-source agent that generates E2E tests for your web app

github.com
8 points·by oceanparkway·2 anni fa·0 comments

Ask HN: Who's self-hosting Supabase and what's your setup?

2 points·by oceanparkway·2 anni fa·0 comments

Show HN: Fixing ChatGPT's losing-focus bug

chatgpt-ui-fix.vercel.app
2 points·by oceanparkway·3 anni fa·2 comments

comments

oceanparkway
·7 mesi fa·discuss
TLDR; monopolism or quasi monopolism is holding back innovation at a country scale notable enough to begin changing the two century long GDP trend of 2%?
oceanparkway
·10 mesi fa·discuss
This article mentions Cerebras, I tried it out and was 1) disappointed and 2) they started a subscription but gave me no way of cancelling, their billing page is broken :/
oceanparkway
·11 mesi fa·discuss
Brutal
oceanparkway
·12 mesi fa·discuss
We need more of this
oceanparkway
·12 mesi fa·discuss
I think the "math" on reliability-over-steps will end up differently than described here in the long term because getting new factual input from the real world should improve the reliability of the end state, and we have all observed agentic systems at this point producing that behavior at least sometimes (e.g., a test failure prompts claude code to refactor correctly).

Whether or not one term in this equation currently compounds faster is a good question, or under what circumstances, etc., but presenting agentic abilities as always flawed thinking resulting in impossible long term task execution isn't right. Humans are flawed and require long, drawn out multi task thinking to get correct answers, and interacting with and getting feedback from the world outside the mind during a task execution process typically raises the chance of the correct answer being spit out in the end.

I'd agree that the agentic math isn't great at the moment, but if it's possible to reduce hallucinations or raise the strength and frequency effect of real world feedback on the model, you could see this playing out differently perhaps quite soon. There's at least a couple of examples of "we're already there".
oceanparkway
·2 anni fa·discuss
I would ask them what their ideal setup is and then compare feasibility. There's probably a lot of indirections/hoops you could jump through but if your security concerns are being driven by your customers you should probably ask them. If it is the case that you need to access their unencrypted data then at one point or another you're going to have to do it, the question is which possible way would your customers feel happiest about? On-premises contract, storing encrypted + customer-specific decrypt keys with a managed auth service, etc etc
oceanparkway
·2 anni fa·discuss
It’s not just a protocol buffer for concepts though (weak wharf Sapir, lakoff’s ubiquitous metaphors). Language itself is also a concept layer and plasticity and concept development is bidirectional. But (I’m not very versed in the language here re ‘latent space’) I would imagine the forward pass through layers converges towards near-token-matches before output, so you have very similar reason to token/language reasoning even in latent/conceptual reasoning? Like the neurons that nearly only respond to a single token for ex.
oceanparkway
·2 anni fa·discuss
Metal desks!
oceanparkway
·2 anni fa·discuss
The presentation of data was nice (slides with a constant format) and the measured amount of labeling was great
oceanparkway
·2 anni fa·discuss
I don't think engineers (especially non-managers) talk or write nearly enough about personality dynamics at the workplace.
oceanparkway
·2 anni fa·discuss
Definitely burying the lede that Tealok is building an alternative to docker compose. Interesting
oceanparkway
·2 anni fa·discuss
Love this. Appreciate the simplicity and intuitive keyboard shortcuts.
oceanparkway
·2 anni fa·discuss
Well, analytics and now warehousing are PostHog’s product.
oceanparkway
·2 anni fa·discuss
Hilarious/great. How did you think of this?
oceanparkway
·2 anni fa·discuss
Notify your customers if you're an API company and everything 100% goes down for 12hr.
oceanparkway
·3 anni fa·discuss
The DOM allows multiple text nodes of an element.

When you update the React tree with a different/longer paragraph node, it's actually removing the element and replacing it with a new one and the selection state is blown away. This is solved in non-text elements by a very simple diffing algorithm that says, as long as it's not a list with .map (which why is why .map triggers a linter rule warning you to use keys), if it is an element with the same tag (<a ...> is the same as <a ...> but different than <img .../>) then React uses DOM nudging to modify the element as opposed to unmounting it from the DOM and replacing it.

Text nodes work differently in the DOM. If you open up devtools and edit a paragraph while its selected in the browser, you'll see the same effect. Try doing the manipulation with JS, same thing (document.querySelector(a_selector).innerText += 'sdfsdf').

Using Fragments instructs React to append multiple text nodes to the same parent, which means we're simply appending nodes instead of blowing them away, which solves the selection state issue.

All made possible by the fact that browsers just cleanly reconstitute the text fragments back together into display text.

Unfortunately this approach does seem to break the built-in apple screenreader, but then I placed the screenreader on ChatGPT and saw how it already totally was failing. It seems surprising to me that OpenAI wouldn't have addressed this access issue. I guess there may be a whole extensive implementation issue with streaming text and screenreaders in general.