Apple Intelligence, Meta's open-source gambit, and the desktop AI tools changing how we work
Back from IBC in Amsterdam: a model battle royale, a semiconductor power struggle, and why Granola won over a hard-to-impress team without anyone having to sell it.
Fresh off a European summer, I was cast straight back into the depths of AI — this time in video, broadcast and multimedia — at an incredible IBC 2024 in Amsterdam, where CaptionHub was exhibiting alongside its video partners. We saw LLM-enhanced translation for specific use cases (using audio, video and caption data to summarise scenes accurately); met partners at ElevenLabs and Voiseed pushing the text-to-speech envelope; chatted with Speechmatics, who power our enterprise live subtitling; and met the top content producers figuring out how to apply AI intelligently to scale media localisation and accessibility.
Outside the booming media-and-entertainment space, the AI and semiconductor industries are throttling up their high-octane go-to-market and product development. On one side, the titans — Meta, OpenAI, Google, Anthropic and now Cohere — are throwing down in a race that makes the space race look like a friendly jog. On the other, chipmakers run headfirst into the future to provide the processing muscle. It's not all smooth sailing: regulatory speed bumps, the reality of AI at scale, the devilled detail of chip manufacturing, and the occasional algorithm-generated cocktail no one asked for.
Apple Intelligence, Meta's open source, and Granola
At Apple's 2024 event, Apple Intelligence took centre stage — a privacy-focused layer that summarises emails, prioritises notifications and understands sloppy Siri commands. Meta, far from quietly ceding the spotlight, doubled down on open source with Llama 3.1, hoping to lead by giving developers the tools to customise their own systems. Think Linux in the '90s, but with more parameters and far less grunge.
OpenAI isn't sitting still — GPT-4o is keeping its spot in the top tier — but Google's Gemini 1.5 overtook it in community benchmarks just hours into August. Personally, having used GPT-4o and Claude side-by-side all summer, GPT-4o has become a daily essential. And let's not forget Anthropic's Claude 3.5 — apparently we're all naming AI like eccentric pets. Mistral popped up with Large 2, continuing its multilingual flex, while new entrants like Black Forest Labs are raising $100M at a $1B valuation. The prize? Enterprise integration, baby.
Through that process, I just became convinced that LLMs were going to change the tools we use for work. It's especially powerful when it comes to spoken language and making that useful.
At CaptionHub it didn't take much convincing to get the whole team using Granola — one of those products that comes by every few years and makes you wonder how you ever worked without it. The UX simplicity is off the charts, and we're a hard-to-impress team. Change management at its best is when you don't need to prove the benefit beyond using the thing. And there's more hardware, too: Plaud's $169 ChatGPT-powered NotePin is making waves in recording and transcription.
The semiconductor power struggle
Nvidia, the undisputed king of AI chips for now, hit a rough patch in August — that Blackwell substrate flaw sent ripples through the supply chain. Like any reigning champ it adapted, switching to CoWoS-S, even if that's older tech lacking a bit of future-proofing. Intel rolled out its 5th Gen Xeon 6 chips for AI data centres and took a swing at mobile AI with Core Ultra, because AI isn't just for the server farm anymore. AI chip sales are expected to hit $50 billion by year-end, and the Fed's 0.5% cut has stimulated heavily financed ticket items like EVs, which helps chip demand.
Enterprise AI's big-money moment
The money's still flowing fast and furious into enterprise AI. Cohere raised $500M like it's pocket change; Canva acquired Leonardo.ai to jump headfirst into generative AI; and everyone wants a piece, from legal platforms like Harvey to healthcare and translation. But enterprise minds are still figuring out where AI fits and how to deploy it responsibly — a term increasingly overused and misunderstood — in a way that delivers lasting value to expensive problems. Enter Retrieval-Augmented Generation (RAG), which helps models pull in external data to avoid embarrassing missteps. Slowly but surely, enterprise AI is getting smarter and less prone to making things up.
The regulatory tug of war
In Europe, the EU's AI Act is all about managing risks and looking tough for the voters — packed with virtue-signalling and high-level guidelines, and a regulatory headache for the likes of Meta and OpenAI. Then there are the US export restrictions on AI chips, which have everyone from Nvidia to SK Hynix watching as China stocks up on high-bandwidth memory. Regulators want to keep AI in check without stifling innovation: one foot on the brake, one on the gas — something's got to give.
AI's quirky side hustles
Just when you thought AI had figured out language, art and the secrets of the universe, it's here to remind us it still struggles with a simple beat — AI still can't make music, at least not the kind you'd put on repeat. And while we wait, it's decided to have a pop at happy hour: AI-generated cocktails and barrel-free 'spirits' are apparently a thing. Maybe this is the Blue Nun of the future. Till the next one.