46 comments

[ 2.9 ms ] story [ 116 ms ] thread
There was a seemingly odd quick sequence of announcements from elevenlabs the last 24 hours, makes me think it's them - notably, I believe they launched 2.0 of their conversational AI today.
The Feynman voice would be great. I've been using it for non-fiction audio books and it works so well.
Does it say "y'all"?
No, it says youse.
Alas, English used to have a perfectly fine 'thou', but then people abandoned it. And now they are re-inventing the same distinction.

Now just wait until people address a single other person with youse, and then have to make up yous'all to address groups.

(Evolution of language is fascinating. I'm just pretending to be upset.)

Ye is really the missing piece.
(comment deleted)
> English used to have a perfectly fine 'thou'

Thou was second-person singular. Y’all is second-person plural.

"You" used to be the second person plural, as a counterpart to "thou".
Exactly. And y'all had to be invented, because 'you' became ambiguous.
A user named eru likes language, I'm not surprised!

Big fan of linguistics and philology myself too.

Edit: Also 'youse guys' is for groups I thought, but maybe you're keeping it to one word contractions, haha

Things I love:

1. Start and stop button. I love this explicit control over who is talking when.

2. Ability to upload files while the voice chat is going. Great idea. Often times I use gpt voice chat for studying, and it's annoying when I need to add another PDF to the context, since I need to stop the chat, upload, and then restart the voice session.

3. Real-time text display during voice chat. I asked you to take the derivative of a function I described, and it outlined its steps, but it wasn't just the transcription of what it was saying.

Things I hate:

1. The transcription is terrible. It took me 10 tries during the conversation to describe f(x) = x^2. Looking back on the transcriptions, it's literally nonsense.

2. There was a buggy moment when the voice conversation started but it was still demoing all the voice options simultaneously. Need some polishing.

Yet, using Abacus.AIs mobile app, you do not need a.. talk.. no talk UI control. It detects when you interject. Would be a nice feature for Claude as well.
But does the bot know not to interject if I pause to think?
Essentially, yes.

From Abacus:

I try to strike a balance. If it seems like you might be gathering your thoughts, I usually wait a bit to give you space, but I'm also here to prompt or help if needed. If you ever feel like I'm jumping in too soon, just let me know!

That's impressive if it's actually able to adjust its own thresholds so that I can tell it to modify its behavoir (rather like Interstellar's Tars being told to adjust its humor and honesty parameters). Without using Abacus I have a hunch that it's just telling you its capable of that. LLM's in general know very little about themselves, only what they're told in the system prompt. But system prompt could include tool use that alters its own behavior.
I thought transcription was a solved problem now. I run whisper at home and it's blazing fast and accurate with the large model <3. If anthropic is much worse they need to up their game. Or just use Whisper until they do.
I like it, but giving Claude a "Deep Research" mode would be better.
Have not used it myself, but Claude has Research mode in beta.
It has Research , works well with Web Search. Saves a lot of time compared to googling and trying to synthesise knowledge yourself.
What a coincidence! I was just added to the "Deep Research" beta.
I really want to like Claude, but I hit their limit WAY too early when I PAID for it, 9 months ago, WAY before I hit any type of limit on gippity. (gippity - gpt , gimminy - gemini).
Haha, I respect calling it gippity. It reminds me of "I call patrick subaru"
I call her gippity, but I abbreviate the name as GPT when typing.

Just like world-wide-web and www.

I really wish Anthropic would focus all of their developer resources on implementing “download all files”.

I know it’s a massive challenge and might take years to get right but the endless copy and paste is wearing me down.

use claude code
I can’t afford it.
Their new MAX 5x plan is flat rate $3/day but IME it's enough to drive all-day multi-concurrent-sessions if you stay on sonnet.

Their MAX 20x is double the cost $~6/day for quadruple the quota.

Keep in mind that Opus chows quota at 5x+ the rate of sonnet.

I have no idea how anyone can go through that many tokens and maintain coherent code. Really, I think I’m missing something I would love to see a video of this being done live. My own experience (since 2022) is having to keep a very close eye on everything that’s happening. Refactoring manually. Going between models. Reformulating the prompt. Etc.
(comment deleted)
(comment deleted)
Having a design doc, implementation and testing plan, strict linter, and strict compiler helps keep the robots on the rails IMHO.

But even then, I never let it git add or git commit, and about half the time it runs in “ask me before you do any edits” mode and re-guiding it in real time as I see things going sideways.

Use Claude Desktop with MCP attached to your IDE (if you’re coding)
You know you could just use the filesystem mcp server and give it access to your project/downloads folder.
From that article:

> According to the report, Anthropic was holding talks with Amazon, the company’s major investor and partner, and voice-focused AI startup ElevenLabs, to possibly drive future voice features for Claude.

> It’s unclear which of those partnerships, if any, came to fruition.

Here's an easy way to confirm that: check Anthropic's "Trust Center" and review any recent updates. https://trust.anthropic.com/updates

Sure enough, on May 29th they have a subprocessor change:

> As of May 29th, 2025, we have added ElevenLabs, which supports text to speech functionality in Claude for Work mobile apps.

I wonder what they're using for speech-to-text?

Maybe also 11L’s Scribe model?
Meh, Anthropic are dead to me until they have structured output.
> Anthropic are dead to me…

They’re dead to me until they fix their over-aggressive auto-ban. Having done nothing more than traveling frequently, rarely using VPN and only using it for coding, I was caught up in a random inexplicable auto-ban. Zero customer service. Appeal process that leads to a black hole. Whatever their technical advances, their user experience when something goes awry is terrible.

LOL wut? It does tool calls, just build a tool that outputs it?
LOL wut? The point of fallbacks are easy resiliency, not coding the same prompt n times through n different providers whaaat
Hn people are too poor to pay for max?
Or some people aren’t seeing the value at $100/mo