Co-Founder at @Phonic_Co. Previously @Stanford CS PhD Dropout, @MosaicML, CS @MIT. I tend to be wrong, but the learning process makes it enjoyable. 🇵🇰🇺🇲
Introducing Phonic-1, the most advanced speech-to-speech model for agentic tasks — from @phonic_co, the speech-to-speech model company.
We’ve had a breakout year and raised a $27.5M Series A led by @firstround with participation from @Lux_Capital, @felicis, and Definition.
"seems like there's a lot of focus on how we can make AI converse like a human, not just have the voice itself sound natural"
@ashlleymo nailed Phonic's mission in a sentence
.@withdavidai has done great work with this new speech-to-speech benchmark.
Really excited that Phonic-1 outperforms leading models from the frontier labs and is #1 on naturalness and content.
Try out Phonic-1 and let us know what you think: phonic.ai/signup
> Phonic-1 [...] replies 1.1 s after the caller stops, against 15.7 s for GPT-Live
Phonic-1 is the speech-to-speech model built for your most challenging conversations.
an underrated concept in duplex speech-to-speech models is time
there's the time dimension of real-time speech input and output (typically represented by tokens that are 20-200ms encoded audio chunks)
then there's the time dimension of reasoning and tool calls
assume 100ms audio chunks for simplicity (moshi uses 80ms, interaction models uses 200ms). this results in 10 tokens/second to be real-time. if reasoning and tool calls operated at the same "clock speed," a 100 token tool call generation would take ten seconds!!
one way around this is fully delegating so that the frontend duplex model can be fully focused on real-time interactions. this is awkward though, as the main frontend model doesn't have context on tool execution and always has to stall when delegating ("hmm, give me a second").
there are many other ways around this, but thinking about time to maintain real-time understanding and generation while also leaving enough time to reason and take action is critical
the best part of launching Phonic-1 has been watching people build real products on it.
if you're a startup going after speech-to-speech, we want to make that easy: 144k minutes, 12 months, and our research + eng team in your corner.
Personally, every technologist has dreamed of this moment. You grow up reading about technology revolutions, but when do you get to be at the true forefront of one?
It's such a joy working with the Phonic team to build the most delightful voice experiences.
AI is using more and more of the world’s energy. But how much economic value are we getting from AI?
We’re introducing Joules-to-GDP (J2G), a framework that measures how well AI turns energy into economic value and shows which tasks, models, and hardware give the best return.
Across 44 occupations in finance, engineering, law, administrative work, and more in @OpenAI's GDPval, we arrive at three findings:
1) The tasks with the highest economic value are not the tasks with the highest economic value per joule. For example, CAD design earns ~10% more per task than building websites, but the latter earns ~27% more value per joule.
2) The most powerful model isn’t always the best use of energy. Gemma 4 31B delivers 56% of GLM 5.2’s net value with 23% of the energy, for 2.5x the value per joule.
3) More chips don’t guarantee proportionate returns. Moving the same model from 1 to 8 NVIDIA H100s uses 3.2x the energy but cuts completion time by only 22%, with about the same value.
With @SnorkelAI, we’re also releasing GDPWorks, a benchmark for economically valuable tasks across 36 additional occupations.
Joint work co-led by @amirzeinaalii and @Avanika15, with @Mersad_Abbasi, @Jon_Hartley_, John Hennessy, @HazyResearch, and @Azaliamirh.
More details below 🧵
Across the great companies we’ve partnered with, it’s clear how often the first business hire helped the company achieve its ambitions. It’s a linchpin role and one of the best opportunities in Silicon Valley. Be that person for @moinnadeem and @NikhilKMurthy.