
Key trends shaping practical LLM development today: democratizing AI with Skyglass, LEDITS, and lean tools; automating workflows (Watto AI, Autonomous Finance); customizing experiences (ai-jsx, Talk to Merlin); multimodal and knowledge-synthesis (Decoupage, System Pro); and responsible AI (Privacy.ai, BlindBox), while tackling data, infra, and prompt-engineering challenges.
Architectural maestro mode with GPT-4: I fed 20-year NAS Ruby scripts (rsync + Glacier) and asked to extract EXIF, reverse geocode, and index data per folder. GPT-4 rewrote, added gems, fixed version issue, parallelized IO, added a progress bar and mutex handling; it ran in ~30s and saved a day.
Calorie Coach is an AI-powered calorie tracker you use via SMS at 325-225-6743 (325-CALORIE) or caloriecoach.ai. It asks your age, gender, weight, and goals to tailor daily needs and track progress; just text a quick food description and you’ll get meal-time reminders. Ask for help for a reminder URL.
Today I launched DREAM.page, an AI-first personal publishing platform, and I’m excited to see how the community uses it in these early days. Check out the video walk-through, sign up here, and DM me to move from the wait list into the actual product.
AI-assisted blog images: Boris Johnson pulls out of Britain's leadership race, saying he has enough MPs to advance but far fewer than frontrunner Rishi Sunak. Our process uses GPT-3 and Stable Diffusion to render a distant, long-lens image of a suited man before the Houses of Parliament.
I built and deployed end-to-end DL image recognizer to detect when Kona is doing her business. Using FastAI, I fine-tuned ResNet-18 on 100s of images, tested with a confusion matrix, iterated, and launched a Gradio app on Hugging Face in ~2 hours. DALL-E 2 for dog-poop data is blocked.
Want to add photos to a post? Use the blogging software I’ve long wanted: select the description and click 'stable diffusion' to auto-insert a stable-diffusion tag and handle it for you. Example: photo of a puppy jumping through a muddy pond, nikon d50, 100mm, bokeh. Awesome results—thanks, rainbow tag.
Upgrade your next Zoom with personality using DALL·E 2 AI-generated backgrounds. Explore a wide range of options and you’re sure to find the perfect background to make your meeting stand out.
Sharing early blog tooling: auto handwriting via <handwriting> tags and photoreal framed image previews via <photorealpreview style='wood-frame-landscape'>; automatic portrait/landscape handling. Also demoing Octopress + Disqus, a Ruby-like processor, a 'Test Render' with iPhone 13 Pro, and DALL-E 2 generated visuals; AI writing tools coming.
Exploring DALL·E 2 generations: from a Leica M11 “Ultimate Question” to a glittery Bald Eagle in Audubon style, dog variations, and high-quality 2D pixel art prompts. Expect surprising results—sad clown portraits, Mao/space shuttle, NYC ice age, Venice pursuit, flying saucers, glowing aliens, and robot–human collaborations; wrappers added later.
Here's Ed’s Banksy-style prompts for DALL·E 2: 'New graffiti by Banksy,' 'Anti-environmentalism' and 'Pro-consumerism' in Banksy style, plus 'Political campaign poster for Joe Biden' and 'New McDonalds logo designed by Banksy' (Ronald skeleton). GPT-3 refined prompts like 'Pro-consumerism ... bright colors' and 'Anti-environmentalism ... destruction of nature.'
I’ve just gained access to DALL·E 2 and here are my first generations: two mummies sipping espresso in Italy (50mm lens, high-quality photo) and a lemon gelato heart in a cobblestone Italian courtyard, plus Banksy-stenciled Bernedoodles, Dali-inspired scenes, and more.
GPT-3 is OpenAI’s 175‑billion‑parameter deep‑learning language model (trained on ~570 GB of text) you access via a simple prompt-based API; it runs on a massive 10,000‑GPU supercomputer, with a 350 GB model, Davinci at $0.060/1K tokens, 50–500 ms latency, and prompts that shape outputs—context is king.
Value of VC: money, credibility, pattern-matching; most success comes from you, the team; two archetypes: A (growth, confident money-use) and B (need to prove risks), better to be A; good VCs provide patience, extra funding, intros, credibility; assess VCs by past ups/downs, key customers/hires; consider not raising money (C).
Be on a purposeful mission, work with the best people, and build a beautiful, customer‑focused experience while breaking new ground—whether in product, design, or business model. A startup should test unproven hypotheses, but never privilege the model over market fit, customer acquisition, and operations.
Seattle Half Marathon recap: foggy downtown, the I-90 tunnel felt like a treadmill with a megaphone cue for half-marathoners. Aid stations were everywhere—no water or GU needed. Madison–Interlaken hills were tough but doable; energy of runners carried us. Marching/bluegrass bands, Interlaken beauty, Krispy Kreme finish, and a baby announcement.
Turn baby photos into durable board books with three options: My Custom Story—laminated 4x5" books from ~$27.95 (6x6" $39.95) with Flickr/Picasa uploads; Pintsize Productions—design from scratch on thick paperboard with optional finger tabs; Treasured Ink—10-page 8x8 ($39.95) or 6x6 ($29.95), add pages for $3 each, professional cardstock durability.
Get feedback with Google Forms and a script. In ~5 minutes per survey you’ll get emails and a spreadsheet. Steps: create a form, allow anyone to respond, paste gist script in Script Editor, set to_email, and enable On form submit triggers with immediate notifications; email yourself the form URL.
If you didn’t get into TechStars, pivot: get a job at Google, work 9–5, and dedicate nights and weekends to your startup. Use day-job cash to fund contract dev/design for validation and seed money, and secure IP with a lawyer and employer blessing—don’t let rejection stop you.