Armando Santana

Tristan Harris is a former Google design ethicist and leading voice from Netflix’s The Social Dilemma. He is also co-founder of the Center for Humane Technology, where he advises policymakers, tech leaders, and the public on the risks of AI, algorithmic manipulation, and the global race toward AGI. I suggest you check out his most recent interview with Steven Bartlett, one of my favorite interviewers (first comment below). He explains: ◼️How AI could trigger a global collapse by 2027 if left unchecked ◼️How AI will take 99% of jobs and collapse key industries by 2030 ◼️Why top tech CEOs are quietly meeting to prepare for AI-triggered chaos ◼️How algorithms are hijacking human attention, behavior, and free will ◼️The real reason governments are afraid to regulate OpenAI and Google #AI #AGI

65

7 Comments

We found an undocumented feature in Apple's SDK and built a wicked feature with it 🤫 Check it out in the video 👇

21

1 Comment

Apple is doubling down on its local-first AI strategy. While much of the industry remains focused on massive, cloud-heavy chatbots, Apple researchers have indeed released a pair of models that prioritise extreme efficiency on consumer hardware. The two models, #FastVLM and #MobileCLIP2, have just been released on Hugging Face to showcase how multimodal AI can run in near real-time without needing a massive GPU farm. Keep an eye on how these models get integrated into the upcoming #iOS19 and #macOS16 betas. The speed of FastVLM is expected to be a cornerstone of the new "Visual Intelligence" features Apple is rolling out for the iPhone 17 lineup. 1. FastVLM: The 85x Faster Model The 85x faster claim actually refers specifically to the Time to First Token (TTFT) when compared to traditional models like LLaVA-OneVision at high resolutions. Traditionally, high-resolution images are a nightmare for AI because they create too many visual tokens for the processor to handle quickly. Apple solved this with a new backbone called FastViT-HD. The result generates 4x to 16x fewer tokens than standard vision encoders. This allows it to understand high-res images (like fine text in documents or complex UI screens) while running 85x faster on a standard MacBook Pro or iPhone. Apple has even released a WebGPU-powered demo on Hugging Face that lets you run the 0.5B parameter version directly in a browser like Safari or Chrome. 2. MobileCLIP2: The Lightweight Powerhouse While FastVLM is built for speed and high resolution, MobileCLIP2 is designed to be the "brain" for mobile vision. Efficiency: It is 3.4x smaller than its predecessor while maintaining (or exceeding) the accuracy of much larger models. This model is what powers features like real-time video captioning, Visual Look Up in your photo gallery, and accessibility tools that describe the world for visually impaired users. Because it’s so small, the entire process happens on the device. Your photos and videos never leave your phone to be "analysed" by a server. IMHO this release is a precursor to the massive Siri 2.0 overhaul (internally code-named Campos) expected later this year. By moving these models to Hugging Face, Apple is signaling to developers that they don't need to choose between AI capability and User Privacy. #AppleAI #FastVLM #MobileCLIP2 #OnDeviceAI #HuggingFace #MachineLearning #Siri2 #CleanTech #PrivacyFirst

12

16 Comments

General-purpose LLMs fail at clinical text. That’s not a bug. It’s a design problem. And one open-source model is showing us why specialization matters more than scale. Meet ClinicalBERT. It’s not the flashiest AI model out there. No viral demo. No billion-dollar funding round. But it was trained on over 1.2 billion words of clinical data from 3 million+ patient records, and it’s solving a problem that GPT and Claude simply weren’t designed to solve: understanding the chaotic, abbreviated, deeply nuanced language that doctors actually use. Here’s the thing most people miss about healthcare data. Around 80% of it is unstructured. Physician notes, discharge summaries, radiology reports. Written fast, full of abbreviations, context-dependent shorthand. “SOB” means “shortness of breath,” not what you think. General models get this wrong. ClinicalBERT doesn’t. Why does this matter commercially? Medical billing errors alone cost U.S. healthcare providers over $54 billion annually in denied claims and rework. NLP systems that can actually read clinical notes with precision could recapture a significant chunk of that. We’re talking about a healthcare AI market projected to grow from $37 billion in 2025 to over $500 billion by 2033. The opportunity is staggering. But what’s really interesting is the architecture choice. ClinicalBERT uses DistilBERT, a lighter, faster variant of BERT. It’s not trying to be the biggest model. It’s trying to be the most useful in a specific domain. It can auto-complete clinical notes, extract symptoms from patient records, predict hospital readmissions, and power medical chatbots that actually understand medical context. With 50K+ monthly downloads and published research in Nature Medicine, this isn’t a side project. It’s community-validated, Azure-deployable, and already being fine-tuned into 43 downstream models. We’re entering an era where the most valuable AI won’t be the most general. It’ll be the most specialized. The models that deeply understand one domain (healthcare, legal, financial) will create more value than the ones trying to do everything for everyone. ClinicalBERT is a quiet proof point of that thesis. If you’re building in healthcare AI, this is the pattern to watch.

72

4 Comments

Founders often ask us: what does it really take to build a durable AI company? In enterprise AI, we’re seeing two models emerge again and again: Oil wells (systems of record) & Pipelines (automation layers). 1️⃣ Oil wells become the source of truth for data & associated workflows. We see two entry points: Rip + Replace: When legacy systems are weighed down by tech debt, buyers are motivated to take the risk to replace with new AI-native systems. hyperexponential rebuilt insurance pricing & underwriting from ingestion to submission; Valon collapsed 25+ mortgage servicing systems into one; Vesta redesigned loan origination so tasks could run in parallel. Greenfield: When no system exists, startups capture customers early and grow with them. Rillet started as the first ERP for SMB finance teams, automating manual workflows, and has since expanded into replacing incumbents like NetSuite. Oil wells take longer to drill, but once established, they create deep, durable moats. Owning the system of record unlocks workflows no one else can build and builds switching costs. 2️⃣ Pipelines (automation/orchestration layers) -- These sit on top of existing systems and automate the “glue work” humans do between them. We’ve seen two main patterns: Fragmented Systems: When many entrenched systems coexist, pipelines unify workflows without requiring rip-and-replace. FurtherAI provides agentic workflows for insurance, automating cumbersome processes (submissions, loss runs, compliance) across multiple systems. Human Middleware: When humans are the bridge between systems, pipelines digitize that work. Concourse builds AI agents for finance teams, connecting into several financial systems so teams can query and analyze without manual effort. Sola lets customers record a workflow once and turns it into a live AI agent for tasks like invoice reconciliation. Customers don’t have to choose. Enterprises will often buy both: a new system of record in one area, lightweight automations in another. But for founders, the strategies are distinct but both can be massive. What matters is not trying to do both at once, but knowing which game you’re playing to win. New post by me & Joe Schmidt IV in comments!

172

20 Comments

A Founders Guide to The Long Build™ - Physical AI Thesis Part III This one is for the builders. In this final post, my partner Will Porteous and I along with insights from builders in the trenches and investors who play the long game share our guide for founders building intelligence in the real world. A few core takeaways: • Sim before steel: Runway is precious, model the system behavior • Capital intensity is not the enemy — misaligned capital is • Build complete solutions before you talk platform • Hardware opens the door, software wins the room • Trust is earned through resilience, not hype

10

1 Comment

Today, Ridges (Bittensor subnet 62) launched its product. It competes with Claude / Cursor. It costs $29/mo for 100 PRs. That's $0.29/PR -- while scoring 73-88% SWE-bench Verified and 96.3% Polyglot (using open source models(!)). Ridges is 5-7x cheaper than Claude Code and Codex at with comparable quality on standardized benchmarks. Built on Bittensor for ~$10m in $TAO emissions while Cursor raised from VC at a $29B valuation. So: Ridges is better. And it costs significantly less. And it was BUILT for significantly less investment. That's Bittensor. Try it here: https://ridges.ai

11