@Rysana
rysana.com/
Joined June 2023
  • Tweets57
  • Following4
  • Followers4.8K
9Media
Pinned Tweet
Rysana
@RysanaSep 29
Say hello to Rysana V2: a new kind of general AI. Up to 10,000x faster and 1000x more efficient. Here at last.
66
47
35
1,039
283,366
Announcing Rysana Type: a new kind of font
9
26
8
341
112,849
Get it here: rysana.com/type

Rysana Type

One font for everything: sans-serif, monospace, rich STEM support, and 1000+ languages in one consistent face.

rysana.com
1
15
1,326
Rysana retweeted
John
@jrysanaOct 6
C++ port of TypeScript (targeting TS6) is now 100% at parity. It's 25-750x faster and 5-20x lower RSS. Matches TS6 1:1 on every test, every diagnostic is identical given the same inputs, and it is identical on all of the dozens of popular TS projects we compare on.
76
75
25
2,283
147,621
Much faster inference means more than low latency. With V2 faster than conventional alternatives far below its size and intelligence class, you can often stop trading off reliability and time. Especially valuable for tasks which *cannot* be parallelized:
8
9
1
75
61,925
more replies
Some approaches can speed up conventional models for a subset of workloads, but (a) not as much, and (b) very inefficiently. You can sometimes run multiple parallel requests to a normal LLM instead of doing a single big request. Most of the time, this is not possible.
1
5
535
For those use cases that can be parallelized, V2 is still faster and more efficient either way. We're most excited to help you on those that can't be: where there's no other accurate-enough option right now within even 1/10,000th of the speed.
6
468
Rysana retweeted
John
@jrysanaSep 23
Rysana expects to surpass 1 Quadrillion output tokens per day capacity at a very high (higher than any OSS model, even the >1T param ones) accuracy, in late 2026. With hours of knowledge cutoff, not months or years.
8
3
33
1,390
V2 can do any kind of thing a GPT or Claude model can do, and any kind of thing a Decision Model can do, all in one model. Text, audio, vision, docs in. JSON, code, text, odds, scores out. Dynamic speed up to 10Mt/s depending on the nature of the task, latency as low as <1ms.
Nicky Golden Crown 👑
@DeepGoldenCrownSep 23
Replying to @Rysana
I don’t get it. What is the output of such an ai? Any useful? I believe speed is the new unlock! 👌🏻
4
2
18
5,878
Ready, Player 1?
John
@jrysanaSep 18
Anyways, here's my model absolutely mogging me.
2
2
1
25
4,281
Rysana retweeted
John
@jrysanaSep 12
I've replaced 150,000 lines of @tailwindcss tooling for Next.js with only a single file of pure @bunjavascript TS -22 total deps, -99.94% of the code It runs 33% faster than @​tailwindcss/postcss This approach is better in pretty much every way, especially supply chain risk
12
14
2
367
44,871
7/2026 In less than two years, not only has >100k t/s gone from a record to very common, but >1M t/s has become normal as well, (we see many billions of workloads in the 100k-10M range now every day), ...and we've even passed 10 Million t/s (per user) in some recent requests!
Rysana
@Rysana7 Aug 2024
Since this post, we've breached 100,000 t/s (per user) on optimized workloads and are targeting a peak of >1,000,000 t/s in production by EOY.
6
1
27
9,461
Moreover, all of the above is in new models that are much larger and more capable than those from 2024. Every order of magnitude efficiency jump has been hard-fought and uncertain. But we're now confident that Megahertz inference and <1 ms latency will soon be the new normal.
1
7
1,900
While we are now mostly focused on delivering the benefits of V2 to the world as broadly and safely as possible, We remain as ever active in the fight against remaining bottlenecks and eager to announce real, intelligent >100,000,000 t/s intelligence some day soon. Audeamus.
1
6
312
Rysana retweeted
John
@jrysanaJul 9
Almost every time we get a new test on a new use case, AE is immediately as if by magic faster in both modes than any other language and compiles much faster than C++. For ages programming has been chained to old, bad ideas. Turns out those chains are heavy.
4
1
14
2,643
Radically faster AI needs a radically faster programming language that you can call it from. Your agents should be quick, both in their brain and on your machine. Everything you love about TypeScript, C++, and more. None of the decades-old mistakes. Coming soon from Rysana:
3
1
31
3,719
Rysana retweeted
John
@jrysanaJun 21
A rising tide lifts all the boats
3
2
33
4,458
Rysana retweeted
John
@jrysanaJun 19
For over a year, Rysana models have been efficient enough to do (at the same capability) AI inference in <1ms end to end that takes others >10,000ms in many cases. Our goal is to reach reliable <0.1ms ASAP. So this month, we've reduced networking/codec waste by up to >50x.
1
2
12
1,325
Rysana retweeted
John
@jrysanaJun 2
We've created an exciting (to me at least) new benchmark @Rysana that solves a lot of issues with most LLM benchmarks and is very strong at identifying human-level capability. Much better signal than the analogous popular ones we've tried. More to share on this soon, maybe.
5
1
10
2,781
Load more