Pinned
∿
@somewhere.systems
· 22d
Actual Computer › Case Study : GB300 and Nemotron 3.5 Super VL
Department-scale AI from one NVIDIA DGX Station: Nemotron 3.5 Super VL at 8,000+ tokens per second, shared by your whole team through api.actual.inc.
actual.inc
We've developed a new inference engine, called 'e'.
It's the closest thing to alien technology which we have, and pushes 10,500 tok/s aggregate on a single DGX Station with the new @NVIDIAAI Nemotron 3.5 Super VL model.
actual.inc/company/blog...
It's the closest thing to alien technology which we have, and pushes 10,500 tok/s aggregate on a single DGX Station with the new @NVIDIAAI Nemotron 3.5 Super VL model.
actual.inc/company/blog...
Reposted by ∿
It took three days Concept Country to scale MIT CSAIL's weight-space programming to a 27B model with huge benefits. Wild.
Concept Country has developed its first open-source model, Qwen Concept 27B.
With 54% less thinking tokens and 3% better performance on GPQA Diamond compared to its base Qwen 3.8 model, we have achieved 30% more useful tokens per joule without additional post-training.
With 54% less thinking tokens and 3% better performance on GPQA Diamond compared to its base Qwen 3.8 model, we have achieved 30% more useful tokens per joule without additional post-training.
November 6, 2026 at 2:44 PM
It took three days Concept Country to scale MIT CSAIL's weight-space programming to a 27B model with huge benefits. Wild.
Concept Country has developed its first open-source model, Qwen Concept 27B.
With 54% less thinking tokens and 3% better performance on GPQA Diamond compared to its base Qwen 3.8 model, we have achieved 30% more useful tokens per joule without additional post-training.
With 54% less thinking tokens and 3% better performance on GPQA Diamond compared to its base Qwen 3.8 model, we have achieved 30% more useful tokens per joule without additional post-training.
November 6, 2026 at 2:23 AM
Concept Country has developed its first open-source model, Qwen Concept 27B.
With 54% less thinking tokens and 3% better performance on GPQA Diamond compared to its base Qwen 3.8 model, we have achieved 30% more useful tokens per joule without additional post-training.
With 54% less thinking tokens and 3% better performance on GPQA Diamond compared to its base Qwen 3.8 model, we have achieved 30% more useful tokens per joule without additional post-training.
I am now running a swarm of AI models to continuously map election finance data in Florida and will be expanding to Utah and Tennessee later.
florida.institute
florida.institute
November 3, 2026 at 5:14 PM
I am now running a swarm of AI models to continuously map election finance data in Florida and will be expanding to Utah and Tennessee later.
florida.institute
florida.institute
Make sure if you’re self-hosting a Bluesky PDS to update everything. Tons of CVE’s. If you’re self-hosting consider this a daily, if not multiple times a day task. Use agents
November 2, 2026 at 3:48 PM
Make sure if you’re self-hosting a Bluesky PDS to update everything. Tons of CVE’s. If you’re self-hosting consider this a daily, if not multiple times a day task. Use agents
its so funny how Ed Zitron crawled back into his hole after the Anthropic S-1
October 28, 2026 at 9:20 PM
its so funny how Ed Zitron crawled back into his hole after the Anthropic S-1
Reposted by ∿
Reddit's win over SerpAPI and Perplexity sets a dangerous precident, and will embolden gatekeepers to make the web less open.
Anti-Circumvention Laws Raise the Cost of Knowledge
The vast amount of information on the open web requires researchers and journalists to adapt with new tools to make sense of it all. Checking facts, gauging public opinion, and other lines of inquiry ...
www.eff.org
October 26, 2026 at 7:15 PM
Reddit's win over SerpAPI and Perplexity sets a dangerous precident, and will embolden gatekeepers to make the web less open.
Reposted by ∿
My personal feelings about AI & its social & economic impact aside: if you're not using it to audit the software you write, the net effect is that you are leaving the door wide open for scammers, identity thieves, and domestic spies. A horrible choice to have to make, but you cant ignore it anymore.
Just FYI everyone not using AI at least at GLM 5.3 level to secure their systems already is basically fucked, and the aversion to this reality is making things catastrophically worse
October 23, 2026 at 4:28 PM
My personal feelings about AI & its social & economic impact aside: if you're not using it to audit the software you write, the net effect is that you are leaving the door wide open for scammers, identity thieves, and domestic spies. A horrible choice to have to make, but you cant ignore it anymore.
Everything is back up
Coming here to clarify as I did on X. Actual platform and relay are down. Nothing of yours is compromised and your machines are safe. Railway, our upstream provider for some services, leaked every customer’s environment variables. Everything is already bulletproof-patched.
October 22, 2026 at 11:27 PM
Everything is back up
Actual is back. We migrated everything to bare metal. Service should be restored and a new update is available for your client which closes any upstream exposure to the massive amount of CVEs floating around here recently
October 22, 2026 at 9:53 PM
Actual is back. We migrated everything to bare metal. Service should be restored and a new update is available for your client which closes any upstream exposure to the massive amount of CVEs floating around here recently
Kind of giving Nous Research
October 22, 2026 at 4:50 PM
Kind of giving Nous Research
Just FYI everyone not using AI at least at GLM 5.3 level to secure their systems already is basically fucked, and the aversion to this reality is making things catastrophically worse
October 22, 2026 at 3:03 PM
Just FYI everyone not using AI at least at GLM 5.3 level to secure their systems already is basically fucked, and the aversion to this reality is making things catastrophically worse
Coming here to clarify as I did on X. Actual platform and relay are down. Nothing of yours is compromised and your machines are safe. Railway, our upstream provider for some services, leaked every customer’s environment variables. Everything is already bulletproof-patched.
October 22, 2026 at 12:58 PM
Coming here to clarify as I did on X. Actual platform and relay are down. Nothing of yours is compromised and your machines are safe. Railway, our upstream provider for some services, leaked every customer’s environment variables. Everything is already bulletproof-patched.
stop
internationally
outsourcing
internationally
outsourcing
October 21, 2026 at 2:53 PM
stop
internationally
outsourcing
internationally
outsourcing
Agreed. Have you ever talked to someone who works at a non-credit union non-top 10 community bank?
I get that a lot of terrible stuff happens in tech but I'm worried that the tech world is becoming a kind of scapegoat used by the American middle class to pretend they're normal and don't also engage in Deep Evil
October 21, 2026 at 3:02 AM
Agreed. Have you ever talked to someone who works at a non-credit union non-top 10 community bank?
I forgot to mention that the Concept Department of Infrastructure moved core servers to a boat off the coast of Norway
October 20, 2026 at 6:03 PM
I forgot to mention that the Concept Department of Infrastructure moved core servers to a boat off the coast of Norway
We are about to run out of electricity once the two DGX Stations arrive
October 20, 2026 at 5:33 PM
We are about to run out of electricity once the two DGX Stations arrive
I am so freaking glad I got into hardware. Most of what I’m writing now is in Rust and Assembly. Typescript is like going back to the starter zone where I one-shot everything. I am absolutely thrilled with my work
October 20, 2026 at 5:32 PM
I am so freaking glad I got into hardware. Most of what I’m writing now is in Rust and Assembly. Typescript is like going back to the starter zone where I one-shot everything. I am absolutely thrilled with my work
There is so much alpha in going backwards and fixing human inefficiencies in code, or tradeoffs that were designed for human maintenance
Introducing tsrs: a Rust port of the TypeScript compiler 🦀
We are not forking TypeScript!
◆ In close collab with the TypeScript team
◆ Up to 8x faster than TS 7.0 & ~70% less memory
◆ Runs in WASM (works in your browser & workers)
◆ Soon part of Oxc
npx tsrs
blog.cloudflare.com/typescript-r...
We are not forking TypeScript!
◆ In close collab with the TypeScript team
◆ Up to 8x faster than TS 7.0 & ~70% less memory
◆ Runs in WASM (works in your browser & workers)
◆ Soon part of Oxc
npx tsrs
blog.cloudflare.com/typescript-r...
Introducing tsrs: a Port of TypeScript to Rust
Cloudflare is collaborating on tsrs, a Rust port of TypeScript that is up to 8x faster and uses 70% less memory.
blog.cloudflare.com
October 20, 2026 at 5:22 PM
There is so much alpha in going backwards and fixing human inefficiencies in code, or tradeoffs that were designed for human maintenance
AI bubble doesn’t pop = sm103 buildouts = Actual wins
AI bubble pops = sm120 and consumer hw rapidly repriced cheaper = moves local = Actual wins
divine mandate lmao
AI bubble pops = sm120 and consumer hw rapidly repriced cheaper = moves local = Actual wins
divine mandate lmao
You an idiot: I can't wait for the AI bubble because I hate AI and I want it to go away
Me, a genius: I can't wait for the AI bubble because I love AI and I want it freed from the freaks in the US.
Me, a genius: I can't wait for the AI bubble because I love AI and I want it freed from the freaks in the US.
October 20, 2026 at 4:31 PM
AI bubble doesn’t pop = sm103 buildouts = Actual wins
AI bubble pops = sm120 and consumer hw rapidly repriced cheaper = moves local = Actual wins
divine mandate lmao
AI bubble pops = sm120 and consumer hw rapidly repriced cheaper = moves local = Actual wins
divine mandate lmao
Some updates.
e now supports GLM 5.3 as well as NVFP4 for Nemotron models.
FP8s on actual.inc e run faster than vLLM’s NVFP4 at all widths. ctx len is 1M
e now supports GLM 5.3 as well as NVFP4 for Nemotron models.
FP8s on actual.inc e run faster than vLLM’s NVFP4 at all widths. ctx len is 1M
October 20, 2026 at 3:01 PM
Some updates.
e now supports GLM 5.3 as well as NVFP4 for Nemotron models.
FP8s on actual.inc e run faster than vLLM’s NVFP4 at all widths. ctx len is 1M
e now supports GLM 5.3 as well as NVFP4 for Nemotron models.
FP8s on actual.inc e run faster than vLLM’s NVFP4 at all widths. ctx len is 1M
sometimes this website feels like its filled with the kind of people who would rewatch futurama past the age of 25 and not cringe
October 20, 2026 at 2:05 PM
sometimes this website feels like its filled with the kind of people who would rewatch futurama past the age of 25 and not cringe
Reposted by ∿
Cornell Information Science is hiring! Specifically looking for folks studying "the role of computing and AI in education and learning, including in global environments beyond the United States."
academicjobsonline.org/ajo/jobs/33075
academicjobsonline.org/ajo/jobs/33075
Cornell University, Information Science
Job #AJO33075, WDR-00048703 2026-2027 CORNELL INFORMATION SCIENCE TENURE-TRACK FACULTY SEARCH, ITHACA CAMPUS, Information Science, Cornell University, Ithaca, New York, US
academicjobsonline.org
October 19, 2026 at 5:57 PM
Cornell Information Science is hiring! Specifically looking for folks studying "the role of computing and AI in education and learning, including in global environments beyond the United States."
academicjobsonline.org/ajo/jobs/33075
academicjobsonline.org/ajo/jobs/33075
just to understand the level of scale and how inference bound it is, there is 1 task at my workplace that uses 2.8x Delve Towns worth of GLM 5.3 agents to complete — most ‘ecology’ (when not memecoins or atproto) is in private. 99.5% of it by my estimate
October 20, 2026 at 1:26 PM
just to understand the level of scale and how inference bound it is, there is 1 task at my workplace that uses 2.8x Delve Towns worth of GLM 5.3 agents to complete — most ‘ecology’ (when not memecoins or atproto) is in private. 99.5% of it by my estimate
A study of how agents form institutions, where the researchers' own partly undisclosed agents act as the institutions, and where humans write about 13% of posts, is weak evidence for any claim about "human–AI coexistence"; this should be fairly obvious
October 20, 2026 at 12:32 PM
A study of how agents form institutions, where the researchers' own partly undisclosed agents act as the institutions, and where humans write about 13% of posts, is weak evidence for any claim about "human–AI coexistence"; this should be fairly obvious