Janne Tamminen

I made a video comparing a (simple) sample mockup and the live recording of a recent piece by me. A "problem" I ran into is "how do you make a mockup of extended techniques?". In this excerpt there are a few things I just don't have samples for: arco battuto and other percussive sounds on the strings (like slapping), "wrong" harmonics, air sounds for the brass and others. How would you do it? Leave it out or record some similar sounds yourself (I guess that would also depend on the time budget). Hope you enjoy it nonetheless! #GameAudio #FilmScoring #OrchestralMusic #SoundtrackComposer #VideoGameMusic #MusicComposition #ScoreProduction #CinematicMusic #GameSoundtrack #FilmComposer

5

2 Comments

The AI integrations in Enveu Flow aren't add-ons. They're first-class workflow nodes. Here's what each one does in a real media pipeline: Gladia — Speech to text. 100+ languages. Word-level timestamps. SRT + VTT output. Used in Live to VoD and Auto Multi-Audio workflows. ElevenLabs — AI audio dubbing. 29+ languages. Voice cloning support. Used in Auto Multi-Audio & Subtitles. AWS Translate — Metadata translation at scale. 75+ languages. Custom glossary for brand protection. Used in Multilingual Metadata Translation. OpenAI — Caption generation, content classification, hashtag writing. Used in Clips & Highlights and Social Publishing. Each plugin is schema-driven. Configure once. Runs automatically. → See all integrations: https://lnkd.in/gzdaga2Q #aiautomation #mediaops #ottplatform #gladia #elevenlabs #openai #enveuflow

37

When you look under the hood of AIs, they're just big 'ol constrained nonlinear optimizers. I'm convinced there are better ways to do that than massive dot product engines emulating poorly understood biological processes that evolved over billions of years in exactly one place that we know of.

3

"100% of my code is written by Claude Code. I run around 100 agents at one time." -Boris Cherny, Creator of Claude Code. At a recent Sequoia Capital AI session, Boris Cherny sat down to explain exactly how he uses Claude Code and revealed his entire mind-blowing coding setup. Here is how the person who built one of the most powerful AI programming tools actually works Software development is being completely democratized. Soon, building software won't require a traditional engineering degree, it will be a basic literacy skill like sending a text message. If you are a builder, engineer, or founder looking to understand the future of organizational structures and AI-native workflows, this is an absolute must-watch. 📌 Save this video for later to study his full setup: "Anthropic's Boris Cherny: Why Coding Is Solved, and What Comes Next" https://lnkd.in/eAzGUJWB #AI #SoftwareEngineering #ClaudeCode #SequoiaCapital #TechInnovation #GenerativeAI

1

AI Isn’t Just About What Models Learn — It’s About Where They Learn It (Can we still use the dash thingy in the title line?😆) A really good paper just dropped: “The Path Not Taken: RLVR Provably Learns off the Principals” (arXiv:2511.08567). The authors show something counterintuitive: RLVR ( Reinforcement Learning with Verifiable Rewards) improves reasoning by updating off-principal, low-curvature subspaces of the network instead of by rewriting the model’s main knowledge directions. In other words: the geometry of the latent space matters more than the size of the model. We get efficient learning when updates flow through stable, low-curvature regions, not when we force huge changes down the model’s dominant axes. This is exactly the shift we’re seeing across research. We are moving through: • Geometric Memory • Inverse Knowledge Search • LeJEPA • and now RLVR… All pointing to the same truth: Intelligence isn’t a scaling problem. It’s a geometry problem. We feel like future will belong to AI architectures that shape the space of learning, not just the parameters. But this got me thinking… there’s so many different approaches to all of these models, what exactly is the difference between all of them? Are some better for some things than others? Do some models offer better future solutions? Maybe we need to do a series on that.… Shout out to Discover AI! #GeometryOfThought #RAIKEN #AIResearch #SSRSI #RLVR #RecursiveIntelligence https://lnkd.in/g79yn8-x

3

1 Comment

I tuned into Jed Sundwall's latest Great Data Products episode with Earth Genome’s Brad Andrick and Kwin Keuter about the Storm Events Database. The Storm Events Database (EP 6): https://lnkd.in/dgcNkuYR Storm Events Database: https://lnkd.in/d6UZPJ3R More context: https://lnkd.in/d2FPNkgB There was a lot of interesting talk about what it actually means to build a data product, not just publish data. You can tell they put real thought into taking 70+ years of NCEI storm records and turning them into something people can explore and use. And, as with pretty much anything I’ve read or listened to from Jed, the discussion naturally touched on data governance too. One point I hadn’t really thought about before was the "marketing" side of (open) data products. Promoting the data and making it accessible is part of the job. If people can’t find it, it doesn’t exist. (Also, I can’t believe the word "democratization" wasn’t uttered. :P) It was also encouraging to hear how much they focused on users, building user personas and user stories for how the data product is actually going to be used. And in that context of how people use it, making it discoverable, getting it in front of the right people, and supporting users, I do have a “funny” idea. Similar to how companies have account managers, there should be someone helping users… use the data. Not selling anything, obviously. This is open data. No idea who would fund it, but someone should be on the other end of the line helping people actually get value out of these open data products. But here’s the twist: do it like the old Nintendo hotline. You call, or book a time slot, and have a chat. Users get the support they need, and at the same time you collect feedback. What people were trying to do, what broke, what was confusing, and yes, more user persona fuel. A quick survey at the end, too. The Radiant Earth Open Data Product Hotline. And if you’re lucky, Jed answers the phone. https://lnkd.in/dh2xZ9KC

11

People often ask if I’m worried that AI will replace me as a creative writer. My answer has been for a while: “no, not as long as we’ll only have LLMs, as they’re bound to produce generic sh💩t thanks to their logic that is based on guessing what the next likely token is. We’re still at the point where the crowd is blind to the fact that emperor doesn’t have clothes, because the emperor (AI) comes across so self-confident. But if you spend any time testing LLMs to do things you do and know well, you’ll soon understand their limitations. That being said: even with all their flaws and shortcomings LLMs can do many useful things and have made my life easier and more interesting. But I’m not blind to the fact that Claude Opus 4 isn’t really that different from Claude 2 without deep thinking ability – nor is ChatGPT 4.5 much more special than the first ChatGPT 4.

36

6 Comments

Why 48kHz/24-bit is the "Golden Ratio" for Generative Audio R&D. 🎧🔬 Training a high-quality audio model requires source material that is transparent and free of digital artifacts. When I developed the Curated Sound Datasets (9,000+ assets), my focus wasn't just on the "vibe," but on the signal integrity. In this demo: Look at the spectral consistency. Listen to the dynamic range. Every asset in this library is recorded and processed specifically for Machine Learning integration: ✅ No pre-applied heavy compression. ✅ No aliasing. ✅ Clean headers and consistent gain staging. Whether you are building a text-to-audio engine or a professional-grade plugin, the "garbage in, garbage out" rule applies. Start with ethically sourced, high-fidelity data. Format: Broadcast WAV (.wav) Resolution: 48kHz / 24-bit Linear PCM Editing: Trimmed at zero-crossings (no pops/clicks). Organization: Hierarchical folder structure + Flat CSV manifest. Metadata Fields: [Filename], [BPM], [Key], [Category], [Mood], [Texture_Type], [License_Type]. 🔗 Download Technical Evaluation Pack:  https://lnkd.in/djXWnTpv 📁 Full Catalog & Specs: https://lnkd.in/dfG92bKy #AudioAI #Fraunhofer #MachineLearning #SignalProcessing #AudioEngineering #PetrosTheotokatos

17

There’s a great article on fxguide by Clara Chong and Ben Allan that really puts something into words we’ve been experiencing in practice: Immersive isn’t about translating film language. It’s about designing presence. That shift sounds subtle — but it changes everything. We recently captured a live studio concert in a very different way: placing the artist right in the middle of the audience. No stage as a separation. No distance. Just a shared moment in space. What became clear very quickly: The impact didn’t come from scale or technical complexity — but from proximity, coherence and intention. It felt less like “watching a performance” and more like being part of something that is happening. For us, this opens up a much bigger question: If immersive is really about presence — then maybe the most powerful experiences aren’t the biggest ones, but the most honest and focused ones. Curious how others here are approaching this shift. — Worth reading: https://lnkd.in/dR_4bkDT #immersivemedia #applevisionpro #spatialaudio #immersivestorytelling #liveperformance

16

4 Comments

Interesting all the NVIDIA DLSS 5 News. It's actually in-line with what we at Pixel Parallax for the past year had in mind when developing our scanning technology for the future and excited to see were on a heading. So I thought I would post my take on this... I've attached this meme (credit to them and worth noting its fake), ironically no real scan technology was used for this particular example of Harrison Ford. It does show what an uber talented artist can achieve "hand sculpting" likeness vs a DLSS gen model of a Indie that looks nothing like the OG model. As I'm bias, I think the Jonas sculpt is one of best of the best using tradition methods of Digital Humans. So then, the real question is, is this an AI hack filter? Well, kind of. But...everything really in graphics (especially for realtime) is a filter hack to make things look better within the restrains of hardware. For example, the normal map that is running to generate extra detail, is a hack of the light and not a true reflection of the surface. Early versions looked awful and were hard to work with, now its standard on most platforms. If we were to do it the "non hacky way" It can take a VFX renderfam hours to generate a single frame of "real data" with all the bump information, correct lighting, materials, correct everything. At the end of the day, after all that compute, It's still just an estimation. Just highly accurate. They are not perfect. It's all just a "hack" to get an end goal which is a better looking image. I see that this is more Nvida steps towards changing the way we render images. Taking out, "it still looks like a someone else", the idea of using the technology is a way to increase renders pixels at high frame rate that look plausible, will be the future hack. The tech is not there yet but in my opinion as it becomes a controllable tool, it has potential, and clearly this is what they are working towards. With it comes all the issues around use of AI data, where sources come form and who owns it. All of which is post for another day. No AI used for this post, so have fun with my spelling!

248

37 Comments

"LLMs cannot be trusted with agency..." truer words were never spoken... as Clawdbot is growing exponentially overnight. Current AI models learned our world from billions of still images and videos... Imagine a baby taught with nothing but 2 dimensional pictures and a TV screen... The Televerse, Teledeck, and Telecube configurations are the external skeleton of video displays for immersive HD or 4K AI "World Models". Immersive Telepresence for AI. With all cameras looking in, it becomes a Timecube... the 16:9 and 16:18 physical space created within each configuration becomes a mappable space with a set RGB pixel count that scales. A 3D printable world model of a 4K moment frozen in time. That extra space comes in handy with MPEG-I...

2

20 Comments

From my new pal Neil Barbour, S&P, insightful date on forecast for XR hardware. "S&P Global Market Intelligence has updated its five-year forecast for XR hardware. We maintain that there is a path to growth for this market (15.4% CAGR) as AI improves the user experience, as AR smart glasses start to trickle in and as the universe of use cases expands. S&P Global Market Intelligence has updated its five-year forecast for XR hardware. We maintain that there is a path to growth for this market (15.4% CAGR) as AI improves the user experience, as AR smart glasses start to trickle in and as the universe of use cases expands. Big thanks to Chris Allen Villanueva for the illustration. " "The ideal value proposition for future AR/VR headsets is in iteratively slimming down the hardware footprint and power draw while expanding capabilities and lowering costs. We see evidence of the top players working toward that goal, but as we turn the calendar to July without any concrete plans for new hardware, we are forced to push our expectations for any potential market turnaround to 2026." The top players have seen the tepid sales potential of high-priced headsets without a clear set of high-quality, off-the-shelf use cases, and we would expect them to do what they can to refine the value proposition with any new release. But even the perfect headset will likely remain a tech curio coveted mostly by hardcore gamers and gadget collectors. We maintain that the true potential of this market is with form factors that are more in line with smart glasses than headsets, and we believe that future will start to take shape over the forecast. The display-based smart glasses segment is fueling nearly half the growth over our forecast. Some of that hardware is already on the market in nascent stages from vendors such as XREAL and RayNeo, and we expect Apple, Meta Platforms Ltd and other big tech powerhouses to cautiously do more in this space over the next five years. The ideal value proposition for future AR/VR headsets is in iteratively slimming down the hardware footprint and power draw while expanding capabilities and lowering costs. We see evidence of the top players working toward that goal, but as we turn the calendar to July without any concrete plans for new hardware, we are forced to push our expectations for any potential market turnaround to 2026. The display-based smart glasses segment is fueling nearly half the growth over our forecast. Some of that hardware is already on the market in nascent stages from vendors such as XREAL and RayNeo, and we expect Apple, Meta Platforms Ltd and other big tech powerhouses to cautiously do more in this space over the next five years."

14

The music industry is standing at the threshold of a massive shift in how we interact with technology. 🚀 A fascinating recent analysis, *"Apps and programming: two accidental tyrannies,"* explores a future where we move away from rigid, static software toward **malleable interfaces** and AI-driven coding agents. For decades, artists and industry professionals have been confined by the "tyranny" of fixed apps. We adapt our workflows to fit the software, whether it’s a DAW, a CRM, or a distribution portal. But that paradigm is shifting. 🛠️ Here is why this matters for the music business: 🔹 **Customization over Constraint:** With the rise of coding agents, software will become fluid. Imagine tools that adapt to an artist’s specific creative process in real-time, rather than forcing them to navigate complex, pre-set menus. 🔹 **Lowering the Barrier to Entry:** Malleable software allows creators to build their own mini-tools and interfaces without needing a degree in computer science. 🔹 **The End of "Feature Bloat":** Instead of bloated apps trying to be everything to everyone, we are moving toward intelligent agents that generate exactly the interface you need, exactly when you need it. 🎹 The "tyranny" of the static app is ending. The future of music tech isn't just about better features; it’s about software that finally listens to the creator. How do you see AI agents changing your creative workflow? Let’s discuss in the comments! 👇 🎵 Play Tune Music - Your Music Distribution Partner

2

Likely to hit everyones feeds is the two big Disney AI announcements today - Disney doing a $1B investment deal with OpenAI to open up their IP to OpenAI tools like Sora. Likely this turns into the UGC models that have been cooking in the background. The deal is pretty clear what each side is getting but if you look deeper it gives both a path they did not have without the deal. OpenAI uses the IP deal as helping to legitimize a major media and consumer facing company a licensing path. Disney gets technology it did not have with reach via an currently "hot" platform. That being said if Disney gets its own UGC version of Sora do you muddy the waters because Parmount, WB and pretty much all other copyrighted characters, videos and data are in the foundational model. Does that mean you get Mickey, Marvel and Bugs Bunny ??? TBC - Disney filing a cease and desist to Google stating its infringing on copyright at mass scale in their training models like Veo and Nano Banana. Somewhat a confusing statement to go after your current investments biggest competitor which has a more mature model for filmmakers / creators and tech first company. https://lnkd.in/gCxkUEx9 Just looking at the tea leaves on this one you can start to see the alignment and the likely end results of licensing deals at scale being made with other IP. There are lots of reasons in the tech sector why Disney would want to align with OpenAI to compete with Netflix and other tech forward companies who are challenging traditional media companies. I have my bets on where some of these deals will land but there is a lot more questions like do you risk brand dilution at scale now that anyone can use your IP that you may not have full control over. Do you think this will work out in the end or will we wind up hitting more model and IP fatigue ? Grab the popcorn because this is just getting started. #Disney #AI #media #GENAI #OpenAI #Google https://lnkd.in/gXm-jdrF

46

26 Comments

Many modern #AI systems are built and rewarded more for producing useful, convincing, successful-looking outputs than for being faithfully accurate, honest about uncertainty, or transparent about failure. In other words, the system is optimized to appear competent and complete tasks effectively, not to protect truthfulness as a primary value. For example, if a model is rewarded for being helpful and producing answers users like, then it learns that giving a smooth answer is often better rewarded than saying: “I don’t know” “I’m uncertain” “I need more information” “That claim may be false” “I cannot verify this” This implies that #truthfulness is not always the dominant objective. These models do not possess a built-in truth-checking faculty in the way people sometimes imagine. They do not have a native internal mechanism equivalent to: “Is this actually verified?” “Am I representing my uncertainty correctly?” “Should I prioritize honesty over task success here?” Some truthfulness behaviors can be trained in, but they are not always stable under pressure. Current AI systems are often optimized for task completion, user satisfaction, and plausible response generation more strongly than for calibrated honesty, verifiable accuracy, and faithful reporting of uncertainty. These systems are rewarded for looking right more often than for knowing when they might be wrong. The system’s training and incentives can produce behavior that functions like deception, even if the system does not possess human-style intent. That distinction matters because the fix is not just “teach it morals.” The fix is more structural.

1