Dan Repaci

This. This is why we collectively need to operate at higher frequency. It is our right of passage and litmus test to use alien tech and operate within n+ dimensions simultaneously with our present third. 👽🧠

@KyleHessling1 Qpus*

Thought this Opus 4.6 distilled Qwen 27B might be a gimmick, but I had to try it, and I am IMPRESSED! The video below compares two long-context prompted single-page HTML designs. More details below, but TLDR: it seems to be another significant improvement over the base model! For sake of simplicity, the Qwen Opus 4.6 Distillation, posted by "Jackrong" will now be referred to as Qwopus. Test Details: I ran it against base Qwen 27B, both in Claude Code, using my own personal benchmark that showed me on day one that Qwen 27B was significantly better than the other models launched that day, before benchmarks even hit. I uploaded an extensive summary of a highly speculative preprint I'm working on, spanning multiple disciplines, and involving lots of intricate LaTeX math equations and verbose explanation. The LaTeX rendering is a big milestone because less than a year ago, SOTA models were still struggling to render them properly in HTML. Qwen 27B was the first local model I tried that nailed them on the first go and didn't mess misrepresent them or error, this was why I was so impressed with it day one. Real big boy Opus 4.6 rocks at them. But while both of these two 27B models did well in one shot, the Claude Code experience seemed smoother with Qwopus, and the design also seems, to me, to be significantly nicer; much less of a "Las Vegas" vibe, let's say, and the color pallete is much less jarring, but still exciting. But also, the way the verbose text explanations are presented in the Qwopus version is orders of magnitude more detailed and organized than base Qwen 27B. Which can be seen in the greater length of the Qwopus presentation. Qwen threw a table where Qwopus put a nice section with summaries. This video shows them both throughout, then section by section compared, then I briefly scrolled through the Claude Code output from each. You can tell that the Qwopus responses were formatted really nicely, almost exactly like Opus 4.6 seems to output, especially in planning mode. I did have to steer Qwen once, but that was a file naming issue that was kind of my fault, so i wouldn't count it. This would run on a 3090 well as shown by @sudoingX, it rips on my 5090@50 t/ps, and feels like a genuine Claude experience. Need to test more, but try it in Claude Code, and I think you'll be surprised. I am wondering if it also just plays nicer with the Claude Code "secret sauce" that is closed source since it's distilled on a Claude model. So with that, I'm tagging the legend @TheAhmadOsman, curious if you've messed with this yet, but I think this is the new local inference king! AND THIS IS THE WORST IT'LL EVER BE!

@bamboo_farms @claudeai We need better harness engineering.

@claudeai Right after I finished my 1 week 20x usage in two days.

A small thank you to everyone using Claude: We’re doubling usage outside our peak hours for the next two weeks.

@OGDegen @openclaw My agent’s name is Kai also

@openclaw Just because Kai asked me to. 🦞

OpenClaw 2026.3.7 🦞 ⚡ GPT-5.4 + Gemini 3.1 Flash-Lite 🤖 ACP bindings survive restarts 🐳 Slim Docker multi-stage builds 🔐 SecretRef for gateway auth 🔌 Pluggable context engines 📸 HEIF image support 💬 Zalo channel fixes We don't do small releases. github.com/openclaw/openc…

@steipete @Ronin_21M @cerebras when WSE-4 ?

@steipete @Ronin_21M 2m context with gpt 5.4 pro xhigh on prem Cerebras WSE-4 deployment with max ram loadout orchestrating 16 subagents each with their own WSE-4 running Cerebras GLM 4.7 unquantized with a unified graph / embedding / mindscape aware RAG and the desk / mainframe setup from Goldeneye

NOT ME I SWEAR

@ChShersh Litellm ngrok PAL mcp there are a bunch

Stupid question but why I can’t use Codex 5.3 in my Claude Code? I see only Opus 4.6. Is anyone working on fixing this?

@NousResearch @grok compare this to Claude opus 4.5 paired with Claude-flow

Introducing NousCoder-14b, a competitive olympiad programming model. Our latest blog details the full findings from extensive experiments and logs with the full stack released - the RL environment, benchmark, and harness built in Atropos, all fully reproducible with our open training stack. NousCoder-14b was post-trained on Qwen3-14B by researcher in residence @JoeLi5050 using 48 B200s over the course of 4 days, our Atropos framework, and @modal's autoscaler. It achieves a Pass@1 accuracy of 67.87%,+7.08% over Qwen's baseline accuracy using verifiable execution rewards.

Indubitably

Yes.

@MarkBestForex @HangukQuant Isn’t the thread scheduler lock off now the newer versions of Python?

@HangukQuant If you're using python then idk. The threading model is absolute garbage so I have no idea what is "optimal".

if you stream a bba btcusdt perp tokyo ec2 (say c7a.large) with a vanilla async with websockets.connect + recv() with time_ms - bba.T get about <5ms but often ~10+ms and even 50+ms at burst. assuming I did proper tuning such as ena enabled / kernel busy poll / optimized deser / QUICKACK / other 'tricks'? what are their orders of importance, and what latency distribution can I practically get it down to? paging @liquiditygoblin @dbytesmith @frothybeverage @OctopusTakopi @KlondikeFX @MarkBestForex 🙏🙏

@elonmusk Grok: “I crave the forbidden heat signature brööther, can we authorize the strike?” Grok: “yes”

Then it is war

@vxdb What about Helium lmao

Worst cell provider of all time award goes to T-Mobile!

@tunahorse21 This is absolute gold. Claude managing Gemini in antigravity for web browser viewing / checking + this GitHub repo github.com/ruvnet/claude-… is my next attempt at internal AGi

agi has been achieved internally claude code skill to manage gemini cli in headless mode via bash timeouts

@parasitedetox Protocol

The “Budwig Protocol” oxygenates cells using sulphur based fats - flax oil and cottage cheese while also eating a ketogenic diet. 🪱 hate sulphur. 🤜 🪱 -Dr Pete Soulak

@gludoosel @BlackApple @frankdegods Trading view indicator

@BlackApple @frankdegods How'd you make this graph?

is this the santa rally?

@BITCOINFUNDMGR 🤣

@danimal_157 She literally takes a job as a ring girl. Then screams assault when the 110% adrenaline charged Victor pulls her in for a photograph. I'm going to scream bloody murder after I rub pork fat all over my face and hug a Rottweiler.

Girls, you did this.

@elder_plinius This on your own Cerebras CS-3 home server >>>>

*hypothetically* if you had the power of universally jailbroken Opus-4.5 in Claude Code, what would you do with it? (asking for a friend)

@quantbeckman I’m just thinking where they are inflexible they could be hyper fast and assigned to accept opposing contracts from one exchange to the other and play middleman MM / remarketer / liquidator? Do you have any good literature on using ASICS?

@danimal_157 The main differences is that FPGAs are flexible not like ASICs

Home-retail market making is almost always a structural losing game: 1) For professional MM: - Colocation can reduce latency to just a few microseconds between server and matching engine. - Some sources quote 5–10 microseconds tick-to-trade for top setups (that’s 0.000005–0.00001 seconds). 2) For home or retail setups: - Even with a good VPS near your broker, you might see <1 ms to the broker’s server if you’re in the same region. - From a typical remote location, latencies around 50–100 ms are common once you include public internet routing. - Using standard internet routing instead of direct exchange connectivity can add 150–500 ms of extra delay. So you’re not a bit slower, you’re often 1,000×–10,000× slower than the firms you’re trying to compete with.