@MiaAI_lab Thanks mia!

@dhh @Itsfoss Super snappy on an old surface go 1824 w 128gb.

@MiaAI_lab @ashxhart Yes. Running your flash next and it’s literally a sub buster. Just printed this fan shroud.

Who wants to tell Marc what this is doing to his brand?

Is this a big deal? I trying to figure out how terrified I should be.

@grok my tokens reset yesterday evening and I am already 30% used??? Is anyone else reporting token glitches?

@Blackwellboy I’ve got a 4” fan under mine. I’ve noticed when sandboxing and there is an inference issue it can spike significantly. I may build a watcher to protect from a runaway KV glitch or something.

@JensenHuang @nvidia Thank you. 🙏

@LottoLabs Is 2600 good for a used 5090?

@nicksortor @SouthwestAir I’m ALP and have no idea what you’re talking about. The new seat assignment works flawlessly. They are also starting to add Starlink to the fleet.

@nikitabier @grok wtf is this?

For parallel tool calls I have found nemo 120 to be the best. It takes some harness work and is not as fast as a sub 36b model but the quality of information and agentic skills can't be beat as far as I have experienced. Performing a long-context validation at ~128k tokens for Nemotron production nemo-direct-test profile and all validated guardrails now. Will be posting a bunch of details soon.

@llm_wizard @NVIDIAAI @NVIDIAAIDev

- Tool calling — works perfectly when instructed explicitly; the first test's
phrasing was ambiguous. finish_reason: tool_calls, correct JSON args. ✅
- Safety — works correctly with a system prompt. Without one it over-refuses.
Hermes always sends a system prompt (SOUL.md), so this is fine in prodution following — thinking mode is consuming all available tokens
before producing output on simple tasks. content: None with 800 tokens burned
on reasoning. Needs enable_thinking: false for short-output tasks. ⚠️

---
Here's the full scorecard:

@llm_wizard @NVIDIAAI @NVIDIAAIDev Thank you Chris. Went from 15 tps on nemo 120 to 55 tps.

@AlexFinn I’m running nemo 120 and loving it. @NVIDIAAIDev

Getting ready. Documenting everything I can to help others. @NVIDIAAIDev

DGX Spark arriving Thursday. Prepping the stack.

Claude built out local push-to-talk voice dictation — no cloud, no API, no subscription.
faster-whisper (small, INT8) on a GTX 1660 Super. Double-tap round, ~1GB VRAM. Runs alongside Ollama + Open WebUI with headroom to spare. Nothing leaves the machine.

github.com/oliviermary/My…

Running on a GTX 1660 Super with CUDA 13.0. Ollama serves local models (Gemma 4, Qwen 2.5, Phi-3) through Open WebUI, all GPU-accelerated on 6GB VRAM. Fast forward to output step. I kept it real time for perspective.

@AlexFinn @pashmerepat This is all your fault. Would’ve got a Max studio but they’re out five months.

@openclaw Apple is now 4-5 months out on M3 Ultra. So will be going the DGX route. Looking forward to the continued support of @nvidia refining the stack for local inference.

Blocked by my hero. Truly a sad day.

The beautiful thing about OpenClaw is that you are training an employee. After two weeks we know how to talk to each other. As an example, a simple prompt could generate an entirely new function to the Outlook add-in. Getting here wasn't a cake walk, but now that the custom features are 10 minutes max and I'm not a programmer (although I'm very computer savvy.) It's shocking to me that Microsoft's Outlook product offering is a summarize button which produces a word salad of the original. OpenClaw is going to put tremendous pressure on the giants to innovate. (the above were fake drafts to protect private data)

Don't like the first draft? Hit "Re-draft" with new instructions like "make it shorter" or "ask about their timeline." Same email context, fresh output. ~$0.001 per draft.

we analyzed 3,600+ of my sent emails and built a style profile. Greeting patterns, paragraph length, how I sign off, even when I use "LMK" vs full sentences. Now the drafts sound like me.

Phase 3 is where it got interesting. We added a "Respond" button. Click it, and the AI drafts a reply based on the email content — including the full thread context. But generic AI replies sound like… generic AI. So...

Phase 2 was opportunity entry. I sell capex equipment to industry. When an RFQ email comes in, one click captures it, parses the data, and kicks off our sales workflow. That alone saved me 15 min per opportunity.

Phase 1 was just syncing. Click a button, the current email gets sent to a local server. Subject, body, recipients — basic plumbing. Nothing fancy but it opened the door.

@KristiNoem When did she become radicalized?

@BillMelugin_ Seems like a wonderful person. When was she radicalized?

@CollinRugg @IfindRetards where are you?

@Axiom_Space What’s this?

@Spotify please demote the project manager who oversaw this release. He needs more time.

@joinautopilot launched two subs on autopilot this morning. What happens when the managers take profits and move to cash? Do the buy dollar ETF?

@elonmusk What is this?

I really feel for @ouraring A device that sells for $180 (with likely higher COGS) then has a battery failure after a few months of use. They are sending me a B-Stock or refurbished unit after a simple AI chat which is nice. Not thrilled to get a used replacement though.

@elonmusk Cancel culture is the worst but trans indoctrination is killing our youth.

Enrich your life.

@Jordan_W_Taylor I highly recommend muting annoying words. Casts a wide net but so worth it.

⁦@BabaRamDass⁩ via Imagine.

@jack @WSTRNAPE Assuming this is a copy. Can you send me a testflight invite?

Makeup in the extreme.

@EmmanuelMacron This explains everything.

@libsoftiktok Reality

@nicksortor Is this Hillary’s testicle lock box?

@Support It seems the target zone for this x is one pixel. Iphone 16 pro. Can you fix this?

@ILAttyGeneral says that the public seeing criminal illegal aliens being arrested in Chicago is not cool. @DrPhil would you please interview the AG. I’d like to hear his side of the story.

@AmericanAir not to be that guy but it's not twitter anymore.

@RealJamesWoods She makes a great custom watch face.

@TONYxTWO She makes a great custom watch face.

@elonmusk My fav

@realBlagojevich Didn’t realize he already did. Can you run again?

@greg16676935420 You may be right. Here is what Grok thinks:

Thanks @Apple for updating the calculator app.

This is a major upgrade in the @Spotify app. A pop up for off device speakers that gives the volume buttons two functions. It may be @amazon but either way nice job!

@VivekGRamaswamy Oxymoron according to Grok.

@ericfisher Let’s see how this ages. Credit to Jack, but old twitter is not what the world wants.

@johnkrausphotos Just ordered John. My son is a freshman nuke e w aerospace minor. He’s going to design inter planetary propulsion systems.

Drudge ages like spoiled milk

More drudge lies by headline.

More lies from Drudge headlines.

Never deleting this app.

Drudge is where to go for blob think. Look at the headlines.

Fucking stud

@clwphoto1 Are these spot welds or rivets?

@thedirtGR That dude is peeing.

Einstein having lunch with Darwin, Dall-E through copilot.

This is a great question. Replies on.

@POTUS Good timing.

@KamalaHarris Good timing.

Meet my attorney.

After a few minutes pull from grill and suffer 3rd degree burns, it’s worth it.

After 15 min remove from insulated baking sheet and place directly on grill.

Here we have nan bread and olives from aldi, tomatoes hot pepper, basil, and dill from the garden.

@TheRoaringKitty Why no Schuab? You broke it kitty.

@wallstreetbets Why down Schwab?

@pulte @ryancohen Schuab still down. This is bullshit, and bullish.

Attempt at capturing some outstanding cloud lighting failed.

@DRUDGE is the new @CNN as proof I offer the current headline. Watch the actual footage, totally out of context. The Blob is in full tilt @MikeBenzCyber