Founder/Chief Scientist @Positron_AI Haskell, category theory, AI, and safety. https://xitter.cf/t.co/OduL9QvPz7 https://xitter.cf/t.co/XwbnuaOUzq https://xitter.cf/t.co/wYaTDISbEB
Here, Now
Joined October 2008
- Tweets13.6K
- Following810
- Followers17.1K
- Likes21.7K
Pinned Tweet
Given a choice between two things, I always try to do the one that nobody will believe me about later.
Edward Kmett retweeted
Thrilled to announce the general availability of
@positron_ai's Atlas Inference Appliance on
@OracleCloud, making Positron the first AI chip startup to be deployed at scale with a hyperscaler. Look out for more announcements this week at Oracle Cloud World! blogs.oracle.com/cloud-infra…
Edward Kmett retweeted
Very happy to be working with @FPuklowski and @GoodisonJason at @general_compute for massive scale disaggregated compute at leading speeds and the best TCO!
Our heterogeneous cloud just got better. We're announcing a cluster scale deployment of @positron_ai as our next partner.
We believe demand for fast inference will outstrip supply for the next decade. Our bet on Positron is another step towards getting fast inference for opensource models into builders' hands.
Hardware goes live Q1 2027.
Thanks to @mitesh711, @trsohmers and the whole Positron team! We're excited for this partnership.
LFG!
We at @positron_ai are spending more on tokens than payroll. We're hiring. Help us fix that. ;)
youtube.com/watch?v=Bzqf25xo…
Edward Kmett retweeted
Delighted that "Flow Typing: A New Lens on Linearity" has been (conditionally) accepted for POPL 2027. Fantastic work by Sky Wilshaw! tinyurl.com/2j26dyyn
Stretching the Storage Manager on the JVM
comonad.com/reader/2026/stre…
Edward Kmett retweeted
A new @HaskelInterlude episode is out, this time with @kmett!
We talked about design principles for Haskell libraries, category theory, cache-oblivious algorithms and his recent work involving Haskell, C++, and FPGAs.
Tune in!
haskell.foundation/podcast/8…
Edward Kmett retweeted
My favorite arcade game, Discs of Tron, running on the OpenAI DevDay edition @modretro Chromatic… implemented fully in the reprogrammable logic of the FPGA, done 100% one shot by Codex GPT 6 Astra. Thanks @pvncher and @PalmerLuckey!
By popular demand, here's an actual screenshot from the Turbo Haskell editor, which should be available using `thc edit` soon. @dysinger's joke spoiled my surprise. ;)
Turbo Haskell (public release announcement)
comonad.com/reader/2026/turb…
GHC running on THC can now run --make, including external-iserv and native linking, and obtains byte-identical .hi, .o and executable files!
It also now supports native-image and AOT compilation, making it a real compiler and not just a JIT.
Next stop is a full bootstrap as the chosen Haskell compiler for compiling its own driver, interop API, and GHC-side tooling through cabal.
github.com/ekmett/thc (irc.libera.chat ##thc)
THC can now use Java's CompressedOOPs to get 32 bit pointers in a 16GB heap, giving it some nice space savings.
It also now can optionally use Project Loom to get us something that more closely resembles GHC's threading model with HECs and virtual threads. This greatly improves MVar performance.
It can also now load and run GHC as a library inside itself with some (gradually loosening) restrictions to do type inference and checking! This is forcing me to work around some very scarily large core call-graphs, a need which is clashing with competing JVM limitations on function size, but that work is paying dividends elsewhere.
Kotlin finally had to get dropped from the runtime. It was causing too many accidental inclusions in the truffle call graph from little things like extra null checks and the like, and that made it no longer pay its way. (I seems like I was the only person on the entire internet using Kotlin for Truffle dev anyways it seems. I missed the memo.) 19% more total lines, and being split across 180% more files hurts, but silently doing the wrong thing hurts more.
github.com/ekmet/thc (irc.libera.chat ##thc)
THC now covers at least nominally all ~1500 or so of GHC 9.14.1’s primops, handles GHC’s full traced callstack and async exceptions, and can load and run real applications using cabal: e.g. hscolour, happy, alex, doctest, etc.
I’m very close on running pandoc (a minor wibble about mixing native and managed libraries and lack of rtsopt pass thru kept me from declaring victory) and can just barely load GHC itself as a library with a small source modification (due to limitations around signal handling and the JVM), so I have a path to a REPL.
Need to fix up disparities between the bytecode and AST backends to fully declare async exceptions resolved and both as supported.
And I still want native-image and/or crac support for pre-canning executables.
github.com/ekmett/thc
THC now loads Pandoc cleanly and it actually works. I've begun the long road of getting intermediate build artifacts under control, and fixing performance after this massive capability expansion pass.
So far so good. The 5.4GB of core and debug data that executing it required (from all the transitive modules, etc.) at first is now down to ~60MB between replacing json with a custom binary format and zipping the result. That hurts a lot less to load!
And thus the Core Binary Distribution format was born.
For the curious, Turbo Haskell development is proceeding on IRC in the ##thc channel on irc.libera.chat
Feel free to join!
THC just successfully imported Control.Lens and ran code directly from it... before it was capable of running "Hello World".
My priorities may be a _little_ skewed.
github.com/ekmett/thc/ #Haskell
THC just successfully imported Control.Lens and ran code directly from it... before it was capable of running "Hello World".
My priorities may be a _little_ skewed.
github.com/ekmett/thc/ #Haskell
I've resurrected the comonad.com/reader, interleaved a ton of contemporaneous content I had stored elsewhere, and made a couple of the articles interactive by moving their demo code into web assembly using modern GHC.
Please let me know if you run into any issues. #Haskell
Edward Kmett retweeted
Thrilled to announce @positron_ai's $875M Series C at a $5B valuation!
This round was co-led by @NEA, @Atreidesmgmt (@GavinSBaker), @valor (@AntonioGracias), @AndraCapital, @SemiAnalysis_ Capital (@dylan522p), and Jim Clark (Founder of SGI and Netscape). An incredible group to have backing what we’re building.
Also very happy to have strategic backing from VentureTech Alliance, Hudson River Trading, Cisco Investments, and Naver Ventures, alongside continued support from @Arm , @jumptrading , @dfjgrowth, Qatar Investment Authority, and many more.
What I’m most excited about is what this lets us do next. We’re deploying more than 50 racks of Atlas at @OracleCloud, and this financing funds our next-generation Asimov chip’s tapeout, the production ramp for Titan, and the manufacturing capacity to meet the demand ahead of us.
The goal remains simple: make running AI dramatically cheaper and more energy efficient.
A huge thank you to the entire Positron team, especially @mitesh711 and @kmett, our customers, partners, and investors. Building hardware takes an enormous amount of work, and I couldn’t be more proud of this team. We have a lot ahead of us, and I can’t wait to share more.
$875M raised at a $5B valuation. Positron in the WSJ.
wsj.com/tech/ai/positron-val…
