Opus 5.5 is incredible at animations tried the circulating prompt on @LiteLLM: "make a dynamic 15-second motion graphics video that shows what an incredible motion designer you are, like it's your showreel for a résumé. go all out."

Gemini 3.8 Flash TTS + Opus 5.5 = Animated explainers! Grandpa explains how rotary phones worked back in the day. Prompted Opus 5.5 to explain rotary phones as a conversation, using Gemini 3.8 Flash TTS via @LiteLLM gateway.

Claude Opus 5.5 is INSANE at recreating animations and video edits! I asked Claude Opus 5.5 to mimic its launch video and it one-shotted this video. Turned it into a skill here: github.com/misbahsy/claude-h… Used it for the announcement of Day 0 support on @LiteLLM

Here's how the full breakdown on my local machine

Jev is a super fast video editor! Edited this entire video in less than a minute, and here's how: First, I transcribed raw vid locally with word timestamps, cut it into 16 beats, then asked Jev these seven questions about every beat: does this line need a visual at all? which card template? (18 to pick from, premade) which text effect? which transition into it? how hard do we punch in on the face? which sound effect? which word gets emphasised? Plus five about the video as a whole. Style, colour, caption treatment, energy, progress bar. 117 questions answered by jev < 1 s ~ $0.0017 Using @typesafeai via @LiteLLM

Jev is definitely insane! What if you could just point and build a website? Uses json-render @ctatedev @LiteLLM gateway to call Jev

Doc-OCR router using Jev @typesafeai A Jev-powered router that looks at a PDF page by page, decides which pages actually need OCR, extracts the rest locally. Result: save cost on # OCR pages + speed

TIL - Noul If you are gonna be using Jev, you might hear this too.

41 common ai slop patterns from ai writing remove it with this one skill. @LiteLLM for multiple reviewers

A gift for your agents... (in case if you missed) github.com/misbahsy/anti-ai-…

anti-ai-slop: a 4 staged skill to remove any AI slop from your codex/cc writing. catches 41 different slop patterns: • ~120 ai words, ~100 phrases • em dashes, emoji bullets, title case headings • watermarks, negative parallelism, colon reveals, throat-clearing openers 4 gates: 1. sanitize invisible chars 2. remove common ai words/chars 3. an agent reviews original vs rewrite, loops back for failed reviews 4. optional: a parallel agent uses a different model to generate additional review + score Works with any of your AI agents.

Embedded Evaluators - The hottest job in tech rn!

Phew… Finally Gemini and XAI can catch up now

For the sake of flykind, I trained a fly brain to find its way out of a room through open doors and windows. It scans the room, detects available exits, navigates around obstacles, and escapes. No more buzzing into walls and closed windows. Now I just need to pass the training to the rest of the fly community.

Asked GPT-6 Astra to help me understand this with visuals, the result... Amazing!

Astra is insanely good at bringing vague ideas to life. Asked it to make a 3D version of a train flying through the Astra galaxy spiral with cosmic bodies in the background. Used a trick to burn fewer tokens; simply asked Astra to check its work frequently. It repeatedly checked intermediate renders to ensure it stayed on track. The model is surprisingly fast. This was on medium reasoning, where it used @threejs and @HyperFrames_ to generate the video.

A tldr video version of the incident (covering first two civilizations) piped.video/rMWQxeRqjvQ

Wait until someone tries to use the terminal inside VS code inside terminal …..

Open-sourcing AuraFarm, a marketing harness for AI agents. (share this with your agents, they'll love it!) Give it a product website and a campaign goal, it generate marketing assets customized for your brand. - It extract's the brand’s colors, typography, imagery, voice, and products - create product-first campaign images, cinematic product ads, UGC , and multi-shot videos - localize for different languages - identify timely campaign opportunities - human approval before posting AuraFarm gives agents the brand context, orchestration, memory, and guardrails needed to use it inside a real marketing workflow. Agents keep campaigns moving. You make the final call. Source: github.com/misbahsy/AuraFarm

One of the best resource to learn how LLMs work and how to train your own version. Open-sourcing an unofficial companion notebook repo, where you can see the code, and run cells in colab notebooks. Repo link: github.com/misbahsy/language…

Who's who in AI! Geoffrey Hinton’s Neural Network of Influence Geoff (@geoffreyhinton) sits at the center of an extraordinary research network. I asked Sol to map how his work, students, postdocs and collaborators connect to breakthroughs like backprop and AlexNet; and to researchers who later shaped @OpenAI , @GoogleDeepMind , @cohere , @Kimi_MoonshotAI and the wider field. Featuring @karpathy, @ilyasut, @ylecun, @aidangomez, @nickfrosst, @RichardSSutton, @hardmaru, @rsalakhu and many more researchers whose work shaped modern AI.

From GPT 2 to Kimi3 - A documentary One video to bring you up to speed on the evolution of LLMs. It starts with tokens, embeddings, the residual stream, attention, KV caching, and next-token prediction; then follows the path through linear attention, the Delta rule, KDA, MLA, sparse expert routing, and Attention Residuals. It highlights important work from @Google, @OpenAI, @Kimi_Moonshot, and many others. Inspired by the article below from @waterloo_intern and the LLM Builder course by @felixrieseberg. Codex + GPT-5.6 for brainstorming, research, and code. What should be the next topic?

What sorcery is this... Made a free PDF text extractor tool. > in-browser, local, and private > 22 pages in < 300ms in this vid > copy text, markdown or as a csv

What if we can understand coding interview problems visually… Here’s the Two Sum problem: piped.video/jty1VPns25A

What the....

How is this free? Give this half a day and learn: - tokenization and embeddings - working of a transformer - model pre-training - supervised fine tuning (SFT) - direct preference optimization (DPO) Walk away with some of the most sought after skills

Added Editorial skill as another skill: - Provide links in prompt -> records a scrolling video using browser for each link - Switches between fullscreen A-rolls and picture-in-picture mode for talking head - Kinetic captions are content aware and dont layer over speaker or visual elements - Automatic thumbnail generation as the first frame Again most of the heavy-lifting is by @HyperFrames_ (as rest of the repo)

Added a new skill for pop-out speaker style and another one for kinetic captions and text-behind-person. @HyperFrames_ remove-background did the trick. Text positioning can be changed based on the video.

New Skill: Raw video + screen recordings → Fully edited shorts for socials And the best part, Claude Code/Codex handle all of the editing for you. Uses @HyperFrames_ talking head as part of the workflow. Repo link below:

It decided what libraries to use and how to use them: Core experience: a GPU-displaced spacetime fabric (64 mass wells + 8 concurrent GW ripples in a custom vertex shader) where you drag any of 25 objects from the palette onto the grid and immediately see curvature, a live geodesic preview of the orbit you're about to create, and auto-circular-orbit assist with real-time a/e/period readouts. Physics (pure-JS engine, node-testable): • Symplectic leapfrog N-body; Paczyński–Wiita black holes (ISCO-like); 2.5PN-inspired GW inspiral for compact binaries; momentum-conserving mergers; BH captures; Roche-limit tidal disruption into rings; wormhole teleportation (labeled SPECULATIVE everywhere) • A hybrid integrator that Kepler-drifts dominated orbits when a frame spans too many orbits — this is what makes the timeline work from real-time to ~100 Myr/s without exploding, and the seamless AU → galactic-core scale transitions • Stellar evolution (main sequence → red giant → planetary nebula/supernova → WD/NS/BH) and planet formation (embryos sweeping dust inside their Hill radius) Game layer: 13 progressively-unlocked challenges (orbits, stable systems, gravity assist, Lagrange, binaries, Mars impact, Roche, TDE, inspiral, planet building, S-stars, galaxy merger) with localStorage persistence and per-object education panels that explicitly flag accepted vs speculative science. Replay mode with timeline scrub and event re-fire. Cinematic slow-mo, camera shake, and fly-to on catastrophes.

Kimi K3 (@Kimi_Moonshot ) built this Space Time Continuum simulator for me and now I can't stop playing. Wanted to test the frontend capabilities of Kimi K3 along with its understanding of physics, so thought of giving it a topic that I enjoy. Built this using Kimi code with around $20 in credits. It also added a game within the app to help learn the concepts in a fun way.

1/What? "Kimi K3 designed a chip to serve a nano model built on its own architecture" "In a single 48-hour autonomous run, K3 built, optimized, and verified the chip using open-source EDA tools on the Nangate 45nm library. Within 4 mm², the chip closes timing at 100 MHz and sustains over 8,700 tokens/s decode throughput in simulation, packing 1.46M standard cells, 0.277 MB of SRAM, and an INT4 MAC array with fused dequantization. A chip built by a model, for a model, reflects K3's long-horizon agentic capabilities."

WHAT?!?! Kimi K3 better that Opus 4.8 on all benchmarks!

What if voice chat was more visual? Gave GPT-realtime-2 access to visual tools and it’s enjoying explaining visually. Built using Sol in codex.

AI slop giveaway... these random icons

Seems like they heard us

Access extended, but no fable tokens remaining for the week…needs limits reset

The kind of things Fable can catch that others might struggle...

Getting warmed up for Fable…

Some folks are tokenmaxxing by running parallel agents and loops > keep a running task list or kanban board for each project > use opensource tools like mngr usage to keep track of usage quota an hour before token refresh > if quota remains, run /goal to loop through remaining tasks > make sure to send a message at the end of refresh window + 1 min, to start a new window right away > ask codex or claude code to automate this as a script nitter.net/MisbahSy/status/206984…

PRO TIP: Do this before you cancel your max subscriptions and switch to GLM-5.2 by @Zai_org. Run a bunch of parallel agents using this open-source tool to compare against GPT-5.5 and Opus 4.8. You could have your own evals based on the tasks you usually do. And then decide based on the results.

Planning to make a community of founders to track the two numbers needed to Earn a Billion. Anyone interested to join? > the metrics visible to the whole community > if someone starts falling off, others help bring them back > investors can contact the consistent winners > community wins shared here on X DM if interested

How come fresh fruit juices are not big in US/CAN? Would love this during summer

One could connect this with a nail art printer and generate endless worldcup themed art #FIFA