Arian Gibson

Innovation Architect. AI & Automation Expert. Oh, and I grow really spicy peppers 🔥🌶️ . #CarolinaReaper #PurpleReign

[Arian Gibson @ariangibson]

@RileyRalmuto Wow - that is a killer UI. 🤯

[Arian Gibson @ariangibson]

@CompleteSkeptic @timo_rf 🔥🔥🔥

[Arian Gibson @ariangibson]

@notkevinzhang Thanks, Kevin! YOLO-ing Jev straight into prod. LEEEEEEROY JEV-KINS!

[Arian Gibson @ariangibson]

@notkevinzhang DM sent!

[Arian Gibson @ariangibson]

@thsottiaux

[Arian Gibson @ariangibson]

@Teknium Once it's done - run it again with Astra! 🔥

[Arian Gibson @ariangibson]

@thsottiaux Your team has proven over and over again that you are committed to make decisions in the best interest of the builder community! Again, this is fantastic news - thanks, @thsottiaux.

[Arian Gibson @ariangibson]

@sama GPT-5.6 Sol was already SO GOOD! Looking forward to working extensively with this new model. I didn't think anything would get me to consider prying Sol out of my workflows anytime soon - but here we are! 😊 Congrats to you and the team!

[Arian Gibson @ariangibson]

Congrats to the team at @OpenAI on the release of GPT-6 Astra! At this point, we're starting to need some more difficult benchmarks. 🤣

[OpenAI @OpenAI]

GPT-6 Astra is state-of-the-art on FrontierMath Tier 4, ARC-AGI 3, and TerminalBench-4.0.

GPT‑6 Astra is also a major advance for scientific discovery, with state-of-the-art performance on Terminal-Bench Science 0.1 and HealthBench Pro.

[Arian Gibson @ariangibson]

Mythos 👀

[Claude @claudeai]

We’re introducing Claude Fable 5.1 and Claude Mythos 5.1.

They're the world’s most advanced models for coding and knowledge work.

[Arian Gibson @ariangibson]

@johnternus Nice work on the new M5 Ultra. 1.2 is no joke! Changed the game.

[Arian Gibson @ariangibson]

@MatthewBerman @bot Free Ultra Plan?!?! Sign me up!

[Arian Gibson @ariangibson]

Fun fact from building a token-optimizing web scraper for AI agents: after you strip a page down to Markdown, inline link URLs are still 40–60% of the remaining tokens.

Firecrawl Lite v1.5.0 moves them to a links[] array instead. Same info, half the tokens. Example Wikipedia article: 58k → 30k chars. Massive token savings!

Try it free: One process, no account required, works with any agent that speaks MCP or the Firecrawl SDK. Using it daily with @NousResearch Hermes Agent.

https://t.co/eRU4TiSAkf

[Arian Gibson @ariangibson]

@elder_plinius Bruh - I blame you. 🤣

https://t.co/X6ZTGrARi7

[Arian Gibson @ariangibson]

@RoundtableSpace Notice the pattern?

Hmmmm... 🤔

[Arian Gibson @ariangibson]

WooHoo! Love it - thank you! 🎉 Literally just hit my weekly limit earlier this afternoon! Night and day experience from all the missteps Anthropic is making with OpenClaw, Oauth, Claude Code neutering, etc. The builder community is 100% paying to attention to how well you have been building good will - capitalize on the exodus! Nice work! Now it's back to building something cool!

[Arian Gibson @ariangibson]

That 2.5GB isn’t the full model footprint—it’s the active working set after aggressive compression.
Between MoE sparsity, pruning, low-rank decomposition, quantization, and likely paging experts in/out, you can get the runtime memory that low.
But the full 26B still has to exist in compressed form somewhere.

[Arian Gibson @ariangibson]

Unfortunately it’s a scam… dead giveaway is the price is TOO low, and the seller has little or no feedback. I got excited about the same thing a few months ago and was thinking hmmm… Money Back Guarantee - what could go wrong! 😑 Almost YOLO’d it - but decided against it. Post how the s*show unfolds… so we know how the scam works. I think they’ll ship you something, but it’s not actually the Mac Studio - hoping I’m wrong! Good luck!

[Arian Gibson @ariangibson]

@MatthewBerman Let's Go! 🔥

[Arian Gibson @ariangibson]

Spend any amount of time running complex agentic workflows and the gap between open and frontier becomes visceral, not academic. Peter's right - "not there yet" is doing a lot of heavy lifting. The gap isn't closing as fast as the benchmarks suggest.

[Peter Steinberger 🦞 @steipete]

@sbaratelli @nvidia @openclaw most folks will want as much intelligence as possible, and open models aren't there yet.