I've shipped another Model Context Protocol, and it kind of shocked me how useful it is.
00:00
gerred
10.2K posts
gerred
@sloppenheimer
principal member of technical staff, building frontier artificial intelligence. chasing the speed of light 💫🌌
maine
Joined March 2020
- Replying to @lefticusParsers and compilers could really stand to be broken out, I think people see a giant daunting compilers course and lose sight of how to write (and use) parsers in more real settings, understand tradeoffs. Lots of little concepts like that bundled up with other dragons.
- everyone at every other ai company should be so worried that google went from everyone yelling about how to even use the site to to Paige endlessly posting what you can do, for free, nearly every day. she's in antarctica and utterly relentless when she should be on vacation.😅 I wanted to see how big these icebergs are, compared to the geographic sizes of American and European cities So this is 1⃣ using Gemini in @GoogleAIStudio with @Google Search grounding to get data, then 2⃣ using Gemini 2.0 to create an app displaying those data in @jsfiddle:
00:00 - Holy shit. Flash 2.0 Thinking Experimental is the first reasoner other than o1 and o1 pro to pass my private bench. It's not 100%, but it suddenly made the actual eval not binary. @JeffDean @OfficialLoganK @DynamicWebPaige wow.
- OK, @huggingface I'd say pulled it off with smolagents and their agent definition. Maybe clients do just need to be a negotiating party and we need to keep investing in the LLMs themselves. I love the immediate clarity of this.
- Hey all! I've been building out a coding agent - more on that tomorrow. Along the way there's been a lot of questions and misunderstanding about how Claude Code, etc. works. So, I went ahead and wrote a book into the differences of how it works vs. other agentic systems.
- I've been having fun over the past few hours and I noticed OpenAI's Codex runner was OSS, so I took a crack at optimizing it, did a post along, and analyzed it. Links in comments. 9GB to 56MB-4.5GB 4 hour build to 4 minutes User-extensible builds Working to be transparent merge
- Replying to @simonwI have a pipeline for taking recipe videos from YouTube, running the video through Gemini to extract ingredients/shopping lists/steps/notable moments (if the chef has a cool tip!) with timestamps at step and ingredient use, note levels. Just tried with 8B and no difference at
- Replying to @ludwigABAPi'm sure linkedin is calculating how many training runs ilya should've already done now at 5.5m a pop 🙄
- hey all! i am thrilled to announce i have joined @sfcompute to cook up the largest contiguous ai superclusters on the planet.
- Replying to @thorstenballHaha I love the honesty of it. “no please I actually need help I suck at this”. honestly makes me more likely to sympathize and help.
- Replying to @__alulayou have been judged and have passed the test. you will diminish, and go into the gnome site, and forever remain human.
