Eggert was by far the hardest professor I had at UCLA Literally the most insane exams. Studying was essentially a waste of time You’d walk out thinking you actually got a zero. Awesome guy though

What a wild experience this was to actually have agents record product videos. Using @nymagent for the rest of the series

I would pay double for my grok @bots to have multiple computers vs. a shared one

With all the reverse-engineer anything and decompilation talk, its helpful to think about why software is purchased vs. built The idea that "all software IP is worthless" brushes over what makes software valuable Software is useful when it produces correct outcomes. And correctness is easy to measure in some platforms (e.g. Photoshop) vs while harder in others (e.g. finite element analysis/materials engineering) Someone can quickly look at a photo or SVG created in Adobe and say "that looks good to me". The underlying operations to get to that output do not need to be verified But for software, like FEA platforms, where the real-world impact of the outputs rely on hard-to-validate internal code, software correctness is priority #1 and unsupervised, reverse-engineered AI software becomes harder to trust It all boils down to trust and accountability Users rely on software that they can trust to behave correctly and reliably. There's always a risk tolerance when choosing software that goes into the ROI calculation So if you're a business procuring new software, are you willing to risk your job or company to move to a new unvetted, seemingly-identical tool? We've all seen AI make silly mistakes, write weird code, delete prod databases, and although those mistakes are declining with each new model, the risk approaches zero but its not there yet And for sake of argument, lets say the software was "100% correct" because AI is now just that good - would you actually fully trust there's no faults hidden in the system? Are you willing to take AI's word that it's true? A user's perception influences behaviors just as much as reality. I do think its obvious that a lot of software is no longer as valuable as it once was, but that doesn't mean all software So what happens then? I think more resources (labor and tokens) will go towards pursuing harder and more impactful problems Some platforms will likely crash and burn entirely bc the outcomes are easily verifiable and twins can pop up overnight, while other products will be pushed to innovate with a more competitive market as the floor raises And the coolest thing is the consumer wins in both

I just created an entire series of product tutorials with Grok @bot + @ElevenLabs MCP. Probably the highest leverage thing I’ve done in ages. Customers always ask for videos on new Jellypod features, best practices, and ways to get started. As a single-person team it’s hard to record a video for everything. Writing the script, recording, editing, etc. takes time especially when the platform improves and changes so frequently. So I had my bots design, record and edit multiple 7-10 episode series of screen-recorded videos with voiceovers, showing exactly the best practices for everything you can do within Jellypod. Exceeded my expectations.

If I want instrumental music, why pay for AI music generation when LLMs can write and render MIDI? And as a pianist, I was curious. So I put Claude, Gemini, and GPT to the test. I built a "music-gen" harness (repo below) that helps models create compact, multi-track scores in MIDI from a prompt. Then I told them: "Create the best original instrumental music you can. Make it express your personality as a language model." The outputs were cool. Opus 5.5 was peaceful yet upbeat. Gemini was full of energy and electronic. And Sol 6.1, idk what to call it. Were they great? Not really. But I was pleasantly impressed both on xhigh and low thinking Open-sourcing the harness so it can be used within your app or by Claude Code/Codex whenever instrumental audio files are needed. Hopeful that hyperframes or videos that just need some basic background music can use this. You can require/guarantee specific durations (i.e. a 30s track or 16 beats at 120bpm) or request patterns in the prompt like a "fully-loopable" output or a song in a specific key. The harness is minimal and generates tracks, keys, and notes token-efficiently by using repeatable patterns (a chorus, verse, chords) that can be represented once and reused versus being re-output multiple times. Then we transpile this into a full MIDI output at the end after the score is fully generated. The MIDI is then run through a synthesizer and sound fonts to bring the audio to life in .WAV file. The renderer is composable, so you can even bring your own SoundFonts or even connect your favorite DAW like Logic Pro or Ableton. github.com/Jellypod-Inc/musi…

Text-to-speech is growing faster than ever but the tooling around AI audio is still super underwhelming. But today celebrating a milestone - SpeechSDK just hit 10k/mo npm downloads! This library is just my small open-source contribution to the audio ecosystem. Backstory: When a new TTS model is released (e.g. last week Gemini 3.8 Flash TTS, this week ElevenLabs v4), it's harder than it should be to swap models. I should pick a model based on its output quality, not because I built business logic around certain provider behaviors. The SpeechSDK is like @vercel's AI SDK for TTS, but specifically-tuned for audio and its edge cases. For example: Some models have native conversational dialogue. Other's don't. The SDK includes native multi-speaker stitching for unsupported models. Some models include word timestamps in the response. Others return character-level granularity and some (cough @googleaidevs cough) none at all. Some support emotive [audio tags] with square brackets. Others with <angled>. In some, this is unsupported and accidentally speak them (we don't want that!). Some default to WAV or PCM, and others MP3. Most models don't even normalize outputs to ensure all voices come out at the same volume. As you dig into each more, you start to find little provider-specific quirks and edge-cases that your business logic depended on, making that TTS model change not an easy swap. And since AI still can't "qualitatively hear" well, some of these bugs often go unnoticed. All these audio-related things are solved in the SpeechSDK. Try every new model with a single line string change and it just works. It's open-source, Apache-2.0, and provider agnostic. Github: github.com/Jellypod-Inc/spee…

Prediction: OpenAI releases a top-level orchestrator to manage all your codex sessions with a single chat input

One thing Gemini 3.8 TTS added and currently flying under the radar is interleaved interjections in multi-speaker convos. Put pipes on disfluencies and interruptions (like |yeah| or |mhm|) and you really get that natural conversational back and forth. @googledevs cooked

Day Zero support in @JellypodAI! Gemini 3.8 Flash TTS is the best model to date to create those incredible podcasts, audio dramas, shorts, etc. So many voices and accents to choose from!! And voice cloning!!

Everyone loves Codex fast mode, but I'd love Codex slow mode if it meant requests were 50% cheaper

Built an algorithm that transforms raster images into "hand-drawn" animations. Great to recreate those whiteboard-style explainer videos without the cost of expensive video models (sorry H3 Max). First, use a model to generate a high-contrast image with minimal visual complexity. Typically white on black (i.e. chalkboard) or black on white (whiteboard). Then, use edge-detection to transform each individual object within the image into pixel-wide skeletons. Trace each skeleton into smooth reveal paths favoring path continuation and directionally compatible fragments. Then, arrange those individual objects to appear based on narrative timing and geometric grouping. To go even further, I designed a set of heuristics to mitigate edge cases like capping path complexity for complicated images and drawing text "accurately". The result? A compact set of reveal paths that follows a somewhat intelligent hand-drawn sequence.

Grok Bok just closed it's first B2B deal for me autonomously, honestly so cool So I created a GTM Bot that has access to our ICP profile data, a basic CSV crm, and an @agentmail inbox. Starting cold, it researched L&D training providers, small business owners, corporate execs, and created a super tight pitch unique to every single person. I started small and scaled up to about 10 a day. Each outbound referenced a recent blog post the lead wrote, a podcast episode they were on, or even a company announcement, and hit each person with a two-sentence cold email using non-corpo jargon speak (this is critical - happy to share more). It followed up at most twice, using the CSV to store research on each individual, and replied autonomously when they responded. Each email was grounded in our company's product offerings and the exact wedge where we could provide the most value for the potential customer. In this case, the lead wanted a demo of @JellypodAI. Grok used our MCP, and actually replied back with a unique product example specifically for them using their content. This led to them taking a call with me, closing the deal, and winning a new customer. All I had to do was show my face on a Google call, answer a few more questions, and boom. Deal closed. Wild

The thing people get wrong here is that garbage TV and movies already coexist today with award-winning films. There’s space for both! This documentary will be so compelling because it explores something that can only be told by humans. And that’s great. At the same time, this has nothing to do with AI or its potential in filmmaking. AI can make soulless slop or accelerate great storytelling. It’s an agnostic creative tool. Could something created with AI reach the same level of emotional connection that viewers will have watching Elizabeth and Nathan? Not at all. It’s a real, human documentary. That’s the point. But generative media enables the next Christopher Nolan, who might be in their bedroom right now, halfway across the world, on a micro budget, who now has the same opportunity to create something that millions of viewers might enjoy. The barriers have shattered. However an artist chooses to express their story, whether that is via film, animation, clay, etc. or AI, their creation will be judged primarily on the quality of their outputs. Not by its means. When someone spends their time watching something, it’s a choice. And it consumes the most finite resource in life - their time. So when a tool like AI democratizes creation and there’s more out there, viewers will slowly demand higher-quality and better content. The floor rises, because anyone can now slap together something. But will people care? Probably not. However in the right hands, it gives true artists a tool to raise the ceiling even higher