Nikhil Singhal

Writing about post-keyboard computing: voice, AI agents, and local-first desktop workflows. Building Vaak, open source.

[Nikhil Singhal @feedbrush]

Most voice tools ask you to move work into their app.

I'm building Vaak so voice works inside the desktop apps you already use.

Open source. Local-first. No Vaak account required. Windows preview out now.

https://t.co/8cGqNxczSN

What frustrates you most about voice typing?

[Arvind Sharma @sarviind]

सड़कों पर गड्ढे, सीवरेज और ड्रेनेज पर वीडियो बनाकर सरकार की पोल खोलोगे तो "आतंकवादी" घोषित करके गिरफ्तार कर लिया जाएगा!

संतोष पंडित की कहानी से सबक लेने का वक्त आ गया है दोस्तों!

[Nikhil Singhal @feedbrush]

1

[Nikhil Singhal @feedbrush]

@amitkilhor It's already there - NPTEL

[Nikhil Singhal @feedbrush]

The more I use AI, the stronger my belief becomes that AI is becoming increasingly intelligent.
But I feel there is hardly the same level of improvement on the creativity side.
I was thinking about the most possible reason for this.
AI is learning from us. And there is enough intelligence in the world that can be fed to AI.
But in terms of creativity, what are we actually providing it?
If you look at the number of creative people around us and the number of intellectual people around us, there seems to be a huge difference.
When I say intelligence, I mean people who are good at maths, logic, solving problems, science, and similar things.
I used to think of this as a left-brain versus right-brain difference — that logical thinking comes more from the left side of the brain and creativity from the right side, and therefore left-handed people might be more creative because the left hand is connected more strongly to the right hemisphere.
But from what I understand now, this is an oversimplification. Creativity and intelligence both involve different parts and networks of the brain, and there is no strong evidence that left-handed people are generally more creative.
Still, when I look around in real life, genuinely creative people seem much rarer.
And that makes me wonder: has human society, over time, somehow suppressed creativity? Or is creativity simply rarer by nature in human beings?
This is only a hypothesis I have been thinking about. I have read a few things here and there, but I am not sure. I am open to other thoughts.
Because when I look at the creative side of AI, I am still not sure how much genuinely creative work it is able to do.
It is definitely becoming better at solving problems.
But whenever we talk about AGI, I usually imagine things like: it will solve the biggest maths problems, tell us something new about black holes, make scientific discoveries, or solve problems that humans have not been able to solve.
But we hardly talk about creativity.
Will AGI become genuinely creative?
Can it come up with something that is truly original?
And I don't even know — do we have proper benchmarks for measuring creativity in AI?
Just a hypothesis I have been thinking about. Open to thoughts. #ai #llms #creativity

[Nikhil Singhal @feedbrush]

@weswinder @sirstripy yes not possible

[Nikhil Singhal @feedbrush]

@weswinder This is a huge claim. I’m using Luna on the max settings, with Fast Mode turned off, and even then my Plus allowance only lasts about a day.

[Nikhil Singhal @feedbrush]

@haider1 But i can se DeepSeek is at $0.06. At this point, everyone is getting cooked by everyone 😭

[Nikhil Singhal @feedbrush]

@suspended_bro Threatening an English teacher is brave—you won’t get punched, but you will get corrected until retirement.

[Nikhil Singhal @feedbrush]

@Adidotdev The real issue isn’t the missed date—it’s the trust gap. Gemini 4 needs public evals for SVGs, tool-calling, and reliability; a bigger number won’t fix that.

[Nikhil Singhal @feedbrush]

@israfill Free access is useful, but the harder comparison is what each model handles well in a real workflow. The cheapest model wins only if verification and context switching do not erase the savings.

[Nikhil Singhal @feedbrush]

@ShiftedMusic “Where the text lands” is the underrated part of dictation. Accuracy matters, but breaking flow to fix or paste the result is what makes it feel slow. How are you handling corrections and app context while keeping everything on-device? @ShiftedMusic

[Nikhil Singhal @feedbrush]

Bad dictation adds steps:

think → speak → open transcript → copy → switch apps → paste → correct

Useful dictation removes them:

think → speak → text appears at the cursor → continue

Accuracy matters. The number of interruptions decides whether voice becomes a habit.

[Nikhil Singhal @feedbrush]

@yuliaisc https://t.co/0TcURtBzCP

[Nikhil Singhal @feedbrush]

One decision in Vaak: local dictation works without a Vaak account.

That makes setup harder. Users bring a provider key; we have to test it, store it securely, and explain failures clearly.

The tradeoff is worth it: the core voice loop does not depend on our backend.

[Nikhil Singhal @feedbrush]

@diegohaz That $1.20/session figure is more useful than a leaderboard. The hidden cost is verification: if Luna needs another model to catch misses, the “cheaper” run may not be cheaper end-to-end.

[Nikhil Singhal @feedbrush]

@0xSero The interesting distinction is what “open” covers here: weights, training data, evals, or all three? The label is doing a lot of work.

[Nikhil Singhal @feedbrush]

@cyrilXBT If agents do almost every task, our remaining job is apparently managing the agents. Production is where that joke gets its second act.

[Nikhil Singhal @feedbrush]

@bindureddy The “inevitable” part is doing a lot of work here. @grok, what evidence would falsify open-source AI dominating by 2027?

[Nikhil Singhal @feedbrush]

@Alibaba_Qwen The benchmark climb is impressive. The less glamorous test is whether the generated app survives a messy repo, mobile layout, and the first contradictory product request.

[Nikhil Singhal @feedbrush]

@DeryaTR_ AI coding tool choice is becoming a pricing spreadsheet with a frontend footnote. The real test is whether the tool saves time after the first working draft.