Ishan Anand
@ianand
Followers
2K
Following
2K
Media
360
Statuses
4K
Demystifying AI at https://t.co/MZrjAFamy5 Prev: VP Product @ EdgioInc, CTO/Cofounder @ Layer0Deploy, MIT EECS
Seattle, WA
Joined July 2007
Wanted to share an AI side project: I’ve implemented GPT2 (an ancestor of ChatGPT) entirely in Excel using standard functions. By using a spreadsheet anyone (even non-developers) can explore and play directly with how a “real” LLM works under the hood. https://t.co/rxNcDnwVe2
8
51
339
CascadiaJS Talk: https://t.co/47csOsj24B Blog post: https://t.co/ee5HVgbnXr Saayn GPT2: https://t.co/LDBbQaEQas Saayn Notebooks:
0
0
1
View Source democratized the web by making every site a learning opportunity. What if we did the same for AI? My @CascadiaJS talk shows how transformers work using view source principles to debug AI right in your browser. Web devs curious about AI: this is your on ramp.
1
0
3
So... how does AI work anyway? @ianand demystifies LLM internals for web devs using a real LLM implemented entirely in vanilla JavaScript that runs in the browser. No Ph.D. needed! Link to our YouTube video in the reply and subscribe to get notified when we ship more videos
1
1
5
We verified the TRM results on the semi private holdout set and they're legit Awesome work and contribution to the open source community by @jm_alexia My notes: * This model is tiny! 7M params, but the rub is that it is relatively expensive to run because pre-training and
Tiny Recursion Model (TRM) results on ARC-AGI - ARC-AGI-1: 40%, $1.76/task - ARC-AGI-2: 6.2%, $2.10/task Thank you to @jm_alexia for contributing TRM, a well written, open source, and thorough research to the community based on the HRM from @makingAGI
13
38
537
Update: I've replotted their data to account for real world token costs and some models change position. https://t.co/zG0yVB7jAQ
Few benchmarks test agents in a real MCP environment. LiveMCP101 does. I replotted the Live MCP 101 data using actual market pricing instead of token counts and some models become more/less efficient when you account for what you actually pay per token. Links below.
0
0
0
Youtube: https://t.co/uryv70Arqo Blog: https://t.co/ly0XuA6G64 Paper:
0
0
0
Few benchmarks test agents in a real MCP environment. LiveMCP101 does. I replotted the Live MCP 101 data using actual market pricing instead of token counts and some models become more/less efficient when you account for what you actually pay per token. Links below.
1
0
1
Interesting comparison from the AI team at Zoom of different LLMs when used as MCP Agents in difficult tasks. Surprisingly they found that open source LLMs didn't get better with increased token usage while closed source models did. Paper: https://t.co/9yzU694OOd
1
2
11
Professor Martin's full essay: https://t.co/nbcP8j6Um4 My blog post:
rogermartin.medium.com
Compelling Desired Customer Action
0
0
0
This gets to the heart of one of the hardest things stakeholders commonly wrestle with. Saying "No." Strategy is inherently about choices and tradeoffs and this test can help cut through strategy theater and force hard strategic decisions.
1
0
0
"A strategy choice is one for which the opposite is not stupid on its face." Example: "Being customer-centric" isn't strategic because ignoring customers is obviously terrible. If every competitor will make the same choice, it's an operating imperative, not a strategy.
1
0
0
The next time someone pitches a strategy, ask: Would the opposite be stupid? This is Roger Martin's (of a "A Plan is Not a Strategy" brilliant litmus test. A short thread. 🧵
1
0
0
Joe G presented on an unknown pdf extraction model from China and Nemo texted the founders “hey a guy here at @AITinkerers is showing your stuff” vibes :)
3
1
7
Web: https://t.co/KJ7dNRlQtB Try It: https://t.co/K1S0OTS2x4 Github: https://t.co/M3yMwQTKLl Discord: https://t.co/g6EO90sIMU YouTube:
0
0
1
New tutorial: Building an AI cover letter workflow with HTML & JavaScript in SAAYN Notebooks We build a smart cover letter workflow from a resume and job description. But the real point is how these AI workflows can be customized with simple web dev skills. Links below 👇
1
0
0
@doomslide @teortaxesTex > this efficiency is paid for by a worse inductive bias than archs that act on a hidden state. Good summary. There's no free serial lunch. Supporting evidence: SSM is supposed to be a competitor to Transformer: as parallel as Transformer in training, but as serial as RNN in
0
1
11
43 years ago today, Knight Rider premiered. Crazy to think the sci-fi I grew up on in the show is real now: 🚗 Self-driving cars 📞 Video calls ⌚ Smartwatch communicators 🤖 Voice AI (with attitude) Happy Birthday, KITT! ...now I just need a holodeck and a warp drive.
0
0
3
Video: https://t.co/946z183VoF Try It: https://t.co/K1S0OTSAmC Github:
0
0
0
What if your AI workflows and templates were as portable as PDFs? Email them to anyone. Run them anywhere. Well now they are. Just posted a new video showing how to build reusable AI prompt templates that work everywhere using SAAYN Notebooks. Video & links in the comments
1
0
1