Jon Morton

@fluorane yes

@_jamorton can you make it better plz

good model

Cartesia’s Sonic-3.5 takes the #1 spot on the Artificial Analysis Speech Arena Leaderboard, surpassing Inworld Realtime TTS 1.5 Max and Google’s Gemini 3.1 Flash TTS Sonic-3.5 is the latest TTS model from @cartesia . It supports 42 languages, including 9 Indian languages, with 500+ voices available out of the box. The model has been highly preferred among voters in the TTS Arena, with its demonstrated naturalness and accurate transcript following. Key takeaways: ➤ Quality: Sonic-3.5 has an Elo score of 1,218 (+16/-16) based on 1,144 arena appearances, placing it ahead of Inworld Realtime TTS 1.5 Max at 1,194 and Gemini 3.1 Flash TTS at 1,209 ➤ Pricing: Sonic-3.5 is priced at $39/1M characters, a premium compared to Gemini 3.1 Flash TTS at $18.3/1M characters, and Inworld Realtime TTS 1.5 Max at $35/1M characters ➤ Speed: 105.5 characters per second, compared to 205 characters per second for Inworld Realtime TTS 1.5 Max and 26.3 characters per second for Gemini 3.1 Flash TTS See more details and listen to samples below 🧵

@_albertgu @cartesia Decent model

Extremely proud of the team @cartesia for launching Sonic 3.5, which sets a new state of the art for TTS I personally led the technical direction of this model; we built it ground up from first principles, and it contains multiple non-trivial ideas that differ substantially from anything we’ve seen in the literature. It’s been very gratifying to see research bets play out and the strong research team at Cartesia continue to grow!

We've raised $100M from Kleiner Perkins, Index Ventures, Lightspeed, and NVIDIA. Today we're introducing Sonic-3 - the state-of-the-art model for realtime conversation. What makes Sonic-3 great: - Breakthrough naturalness - laughter and full emotional range - Lightning fast -

@NikoMcCarty If Rubisco can be made more efficient in a relatively straightforward (for evolution) manner, why is the inefficient version so prevalent?

Rubisco is (arguably) the most abundant protein on Earth. (LPP surely comes close, right?) It’s an enzyme that fixes CO₂ into sugars during photosynthesis. Unfortunately, as most people learn in school, Rubisco is inefficient. Sometimes it confuses O₂ for CO₂ and wastes energy. Plants make up for this in raw concentration; up to half the soluble protein in a leaf is Rubisco. People have been trying to engineer better Rubiscos for many decades, but it's not easy because the proteins are big, do not fold easily (they need chaperone proteins to help out), are made from 16 subunits in land plants. But there's a new paper in Nature Plants that looks really interesting. The TL;DR is that a group in Australia figured out how to express plant Rubiscos (and all SEVEN of their folding chaperones) using a set of 3 plasmids inside of E. coli cells. This enabled them to do "directed evolution" of Rubisco in bacterial cells, and quickly find Rubisco mutants that have higher enzymatic efficiency or that fold better. In addition to the 3 plasmids, the researchers also coaxed E. coli to make ribulose-1,5-biphosphate, or RuBP, which is the 5-carbon sugar that Rubisco smashes into carbon dioxide to make molecules of 3-PGA for central metabolism. Now, the clever bit is that you RANDOMLY MUTATE the three plasmids encoding the Rubisco to make millions of variants. Then, you transform those mutated plasmids into E. coli. If the E. coli do NOT make a functional Rubisco, RuBP levels build up and kill the cell; the molecule becomes toxic. But if the E. coli DO make a functional Rubisco, then they keep the RuBP levels in check and live just fine. Using this "screening assay," the researchers found 46 fast-growing colonies of E. coli. Two of those colonies encoded really useful mutations. One mutation (M116L) makes Rubisco about 25–40% faster. The other (A242V) makes it fold and assemble much more efficiently. They put this mutation into a "hybrid Arabidopsis–tobacco Rubisco," put that into tobacco plants, and measured growth. The plants with M116L grew 75% faster than wildtype. No guarantees this will scale to more useful crops, like wheat and corn and soybeans etc. But it seems like a nice in vitro assay for faster prototyping!

@mickcodez Incredible, congrats!

This time last year I couldn’t do 1 push up, couldn’t jump, couldn’t run. I just did 30 pushups in the driveway after a 30 minute game of basketball. I’m locked in, not just on AI, but being the BEST version of me.

@teortaxesTex @TheHumanoidHub Awesome though that run-time does make you appreciate energy efficiency and energy storage capability of humans.

> 5 feet 5 inches tall — 30 kg weight > Walking speed 2.5 miles/hour (1.12 m/s) — Running speed 7.5 miles/hour (3.35 m/s) > 44 lb (20 kg) carry capacity — 2-4 hours run-time. (H/t @TheHumanoidHub) Those are insanely good specs. Plausible iPhone moment for humanoids.

@teortaxesTex Is Gemini not frontier?

It defies belief that all data center giants including Amazon are still incapable of producing a near-frontier LLM. How long do they imagine this "utility" strategy will work? They must be trying to buy up smaller groups that punch above their weight – magic\.dev, Reka, whatever.

@francoisfleuret Well, I guess it depends on what “commitments” they asked for and if they are related to the previous stated issues (blue check, research data etc) or were commitments around removing content. If the latter, it’s clear EU officials don’t want to say that outright.

@_jamorton I am really asking. The EU does a terrible job at regulating this kind of stuff, but Musk is king at making excessive unfounded statements to paint his "opponents" as corrupted and evil.

Any evidence?

@marikakatanuma Why is it so low?

Kishida..

@aaron_defazio @andersonbcdefg Is it perhaps possible to construct a penalty that can induce sparsity without over regularizing the non zero entries?

@andersonbcdefg Yes, a L1 penalty large enough to induce sparsity tends to greatly over-regularize the remaining non-zero weights.

L1 regularization for sparse solutions - as usually taught - is actually terrible in practice! I’m always surprised how few people know this. To get good results, retrain with the sparsity pattern found from the initial L1 run, but without the regularizer. Works much better.

@francoisfleuret this is why ml frameworks should have been written in javascript, which is a very sane language

2h of debugging. Whatever you say, that's counter intuitive.

@francoisfleuret @intel @AMD It’s the software. Always has been.

What's going on with @intel's Gaudi 2 and @amd's MI300? We hear from supposedly serious sources that they are great and they are nowhere. It's a large scale production issue? Drivers suck and they shine on cherry-picked baselines but are unusable in general? The benchmarks lie?

@marikakatanuma Thanks! 🙏

@_jamorton The context is in the story. In case you need a gift link bloomberg.com/news/articles/…

China's surprise chip breakthrough lifts Huawei's revenue close to $100 billion despite years of US sanctions.. story: bloomberg.com/news/articles/…

@taobanker if you pay me $20/yr i’ll tell you to buy USLM once per quarter

My research budget is mostly finalized, but would be happy to get any last minute recommendations if there's anyone you'd like to vouch for.

@francoisfleuret @clockwk7 @torch torch.compile is not yet good with this sort of recursive stuff anyways, doubt it’d work. natively supported scan like jax is coming soon though! github.com/pytorch/pytorc…

@clockwk7 @torch You are making a very good point, but I am not satisfied! Python 3.11+ not yet supported for torch.compile

Can someone rewrite my pscan.py in cuda triton whatever goes brrrr??? fleuret.org/tmp/pscan.py

@teto8k @taobanker Do you have a source for this information? When he was young his net worth plunged from $55mln to $22mln as he tried for 9 years to make returns from a shitty textile company

young buffet traded 600-700 stocks over a 14 yr time frame Pf churn: 40-50 stocks a year Avg hold period: 6 mnths Cagr: 45-55% ? Coca cola Buffet: 'Hold 20 stocks for lifetime' Cagr: 11% ? Replicate a man in his prime when his account was your size.