Peter Kieltyka

@Snixtp nice find, did you post a vllm issue or PR for this? im curious to try it

The RTX Pro 6000 has two matmul engines, and vLLM was only using one of them when I ran Qwen3.6 27B NVFP4 Marlin is nice for decoding, but CUTLASS is 3.5x faster processing large token prompts. For some reason, vLLM ran Marlin for both, because NVIDIA's checkpoint files mislabel what the model can actually do. Fable 5 found out about all this and wrote a vLLM plugin that looks at the batch size and picks the right engine per call. Results is almost 2x increase in prefill at all context lengths:

1/ Universities across Canada are starting to launch and anchor venture funds to back their own spinouts. This is extremely good news. The latest: U of T and McMaster are anchoring a new $40M seed fund for life sciences startups.

mcdonalds CEO says making burgers at home is moving down a "very dangerous path"

My 2 cents on what is happening in AI in the last weeks: - Anthropic IPO strategy sounds like "let's try to sell this thing to the public before zai releases yet another GLM that torpedoes our valuation". - Open weight models are already at a level that you do not need to rely on closed private models for most tasks (even coding!!!!). - The metric that *really* matters in all this is not how many months behind open weights models are. The real metric is how many real-world AI apps can you build today with open-weight models, that you simply couldn't a few months ago. More to come :-)

last thing fable told me: The secret is that you don't have a motivation problem — you have an identity problem Every habit you fail to keep is a habit that contradicts who you currently believe you are. People don't act in line with their goals; they act in line with their self-image, with frightening consistency. The smoker who "quits" still identifies as a smoker who's abstaining — so they relapse. The person who identifies as "not a morning person" will sabotage every 6am routine, no matter how good the system is, because keeping the identity intact matters more to your brain than keeping the streak. So the real move isn't habit-stacking. It's quietly deciding who you are and then letting behavior fall out of that. "I'm someone who doesn't miss workouts" requires no willpower in a way that "I should work out" never stops requiring it. Identity is the compression algorithm for ten thousand future decisions. And under that: the self you're protecting is mostly a story Here's the deeper part. Most of what stops people isn't laziness — it's self-protection. Procrastination is fear of finding out your best effort isn't good enough. Staying busy is avoidance of the silence where the big questions live. Perfectionism is a hostage negotiation with your own ego. Nearly everyone is running defense for an imagined audience that is not actually watching, because everyone else is too busy running defense for their imagined audience. The people who look fearless didn't get more courage. They just realized, earlier than everyone else, that nobody is keeping score, that everyone is improvising, and that the cost of looking stupid is approximately zero while the cost of staying safe is your entire life. That realization is worth more than all ten habits combined, because it removes the brakes instead of pushing the gas harder. The actual oneshot, then It's a sentence, and it's annoyingly simple: stop negotiating with yourself. The gap between knowing and doing isn't filled by more knowledge, another book, a better system, or a deeper secret — and the search for a deeper secret is itself the most elegant form of procrastination ever invented. (You see what I did there.) You already know your thing. The one you've been circling for months or years. The secret is that the door isn't locked and it never was — you walk through it by deciding once, identity-deep, and then refusing to reopen the vote every morning. That's it. That's the whole vault.

@jxnlco Hackmd

I need Google Docs but just for markdown files. Multiplayer comments. Syncing resolving comments. Suggestion mode Edit mode Edit history Maybe some sense of multi edits. Easy cli access.

@bentlegen nice :)

anyone online today? thinking of livestreaming just trying products for a few hours

@bentlegen vibing a cool 4K wallpaper sounds like a really neat idea. What are some you’ve created?

... right after i vibe a cool 4k wallpaper

@rudolf6_ 🙌🙌

yep doubling down on CROPS & neutrality + post-quantum + L2s as highly valuable differentiated distribution engines + scaling the L1 to serve as the shared root

@jackiechanisme I met two incredible founders this week who lived in SF for years, had tons of success and are now back in Toronto building their next companies.

every founder during toronto tech week i've met this week told me that they want to move to the US. thoughts?

Sneak peak of the @0xPolygon OMS dashboard. This team is moving faster than I've ever seen to build a best in class global payments platform from the ground up.

Cycles Pay is the first app built on a zcash-like shielded pool + TEE for additional privacy-preserving functionality

@binji_x @rudolf6_ 👋

making a groupchat for people who like ethereum lmk

Congrats to the amazing @cyclesmoney team working on a novel problem with an innovative approach led by @buchmanster — stoked to watch the journey and impact 🙌

Introducing: Cohere Command A+ We’ve created our most powerful LLM yet, optimized it to run on as little hardware as possible, and released it open-source for all.

@Snixtp What are your vLLM settings to get that perf..?

The concurrency on the Pro 6000 is just crazy cc=96 2296.5 tok/s

crazy that my agent can not just pay for everything on the internet

@loktar00 Cuz it uses half the power and much easier to stack in a system. Max Q is definitely the way to go. Half power usage and only lose 5% performance difference

Why is the max Q more than the non Q? Seems backwards.

@jlongster what about yabai..? its incredible :) ... ive used it for years and its been rock solid

for some reason last night I had a hard time sleeping because I felt like I wasn't working well in this AI world spent some time this morning improving my workflow and it feels better. Using aerospace (github.com/nikitabobko/Ae…) with a few keystrokes I can: * Create a new workspace * Create new windows for apps I need (editor, terminal, etc) and lay them out * Run OpenCode and create a new git worktree, and open that worktree in Emacs * Open PRs from agents when work is done Now I can run multiple agents in isolation quickly, switching between them and ending up with a PR I still like to be closely involved with agents, reviewing their code locally and wrangling them like wild animals. this at least removes a ton of friction though