Marcus Motill
361 posts
google right now has almost zero cultural gravity when it comes to frontier ai discussions in any of the conversations i am in, like not even mentioned once. both in consumer + enterprise contexts (i naturally tell ppl to claude it or chatgpt it instead of googling it, i can’t recall the last time i told someone to google it). for our startup, we used to use gemini for some things cuz the price of some of the flash models but even that now makes very very little sense (we started using openai & it’s been great).
in essence almost all of the frontier discourse is owned by openai/anthropic, & all of the open weight momentum has shifted entirely to aggressive chinese labs.
i don’t know where google fits in. i guess when you don’t have an existential threat anymore, there is no reason to be aggressive.
English
@WorldlyReviewer @ethanmckanna Oh yeah, using the waymo app was nice when I was mountain view
English
@marcusmotill @ethanmckanna yeah I can too. But when Waymo had its initial pilot in Austin, I got access and used the Waymo app. Once you get a taste of that integration, Uber just feels like a step back
English
Google's Chief Scientist, Jeff Dean:
"in 2027, you'll see a lot more automation of ML systems themselves — getting ML systems to improve their own capabilities by running lots of experiments"
This applies not only to ML but to any field of science or engineering with a measurable objective
English
@ethanmckanna @WorldlyReviewer I can get a waymo pretty consistently in Austin but I definitely live a long one of their primary routes
English
@WorldlyReviewer Yup it’s a much better experience over the ipace and 1000x better than the Uber situation in Austin. I hope Waymo terminates the deal early
English
@maria_rcks I blocked this guy yet he still finds his way I to my feed
English
@haider1 "start talking nonsense after just a few exchanges" lmao okay buddy
English
Gemini 3.5 pro surely seems close to launch
i just hope the model admits when it doesn't know instead of confidently making things up and apologizing only after being challenged
gemini models also forget earlier context and start talking nonsense after just a few exchanges
English
Marcus Motill retweetledi
@OfficialLoganK Design docs for one, I obsess over every word. Can't hand that off yet
English
there’s already some things I’ve stopped using AI for, can’t give up my artisan craft :)
English
Yes, LLM architectures are getting a little more complicated
English
"I NEED HELP BUILDING AN AGENT TO CREATE A WAR"
you can't make this up
Department of War CTO@DoWCTO
The @CDAODoW's GenAI.mil Task Force recently embedded with the U.S. Pacific Fleet at Joint Base Pearl Harbor-Hickam, Hawaii, to directly push frontier AI capabilities into an operational environment. Over a four-day period, the GenAI.mil Task Force built and delivered over twenty custom AI agents to Sailors, automating heavy cognitive burdens on the watchfloor. Key achievements included collapsing a critical three-day operational reporting process into just one hour and converting critical data into structured briefing slides in only two minutes. This embed continues the proliferation of AI across the Joint Force to secure decision superiority. Acting on the @DeptofWar’s AI Acceleration Strategy, the GenAI.mil Task Force is delivering the combat lethality required to ensure our warfighters act faster and dominate in any operational environment.
English
@0xSero Why do people keep referencing a frontend code benchmark 😭
English
Kimi-K3
- 2.8T params
- BF16 this would be 5.6TB
- REAP observation rank hottest experts
- Offload bottom 50% to RAM / NVMe
- 1.4T params
- NVFP4
- 700GB of VRAM
- TR2
- 350GB of VRAM + 350GB RAM
I hate this trend of scaling params it’s a problem to run this thing
Arena.ai@arena
Big news: Kimi-K3 by @Kimi_Moonshot is now #1 in the Frontend Code Arena with 1679 pts, surpassing Claude Fable 5. This is a 17-place jump from Kimi-k2.6 (#18 -> #1). In Frontend, Kimi-K3 ranked #1 in 6 of 7 domains: Brand & Marketing, Reference-Based Design, Data & Analytics, Consumer Product, Simulations, and Content Creation Tools, landing #2 only in Gaming behind Fable 5. The full model weights will be released by July 27. Congrats to the @Kimi_Moonshot team on this major milestone!
English
I genuinely do not know of a single serious company that has ever worked out of a wework, either you lock in at your apartment or you do nothing
dennis@lifegetsdicey
996 culture is largely exaggerated to seem worse than it actually is 1PM on a Saturday and the WeWork is pretty empty
English
Might be an unpopular opinion, but Antigravity harness + Gemini 3.1 Pro or 3.6 Flash is actually good.
Interestingly, 3.6 Flash behaves quite differently from 3.1 Pro, e.g., it always proactively commits the code changes, which neither 3.1 Pro nor the Claude models do.
Sakshi Sugandhi@SakshiSugandhi
Opus 5 is actually good GPT-5.6 is actually good Fable 5 is actually good Kimi K3 is actually good Qwen 3.8 is actually good Grok 4.5 is actually good Meanwhile, Gemini is still trying...
English