The Daily AI Drop - Evening Edition, October 30, 2026

A cheaper DeepSeek, free models with automatic switching, a 120-billion-parameter model running offline on a laptop, a 164-megabyte speech model, and why owning your AI system matters. Hosted by Alex and Jordan.

Listen: Listen to this episode

NEW 10x Cheaper DeepSeek?!

Channel: Mehul Mohan (@mehulmpt)

Link: Watch source video ↗

Mehul looks at a new DeepSeek release claimed to be ten times cheaper, alongside StepFun's Step 5 Preview for agentic work and speech, voice, and video. When a capable model gets that much cheaper, the economics of whole products change: things too expensive to ship suddenly pencil out. Pair cheaper intelligence with more capable agents and the gap between demo and real product gets a lot smaller.

FreeLLMAPI – Access Free AI Models With Automatic Switching

Channel: TheNextNewThingAI (@TheNextNewThingAI)

Link: Watch source video ↗

FreeLLMAPI pulls together free AI models from different providers and automatically switches between them when free credits run out, so work keeps going without a paid plan. The video's example is paying $200 a month for Claude even though not every task needs a top-tier model. For builders watching their AI bills creep up, routing everyday work to free tiers is a concrete way to push costs back down.

Why Own Your Intelligence? Harrison Chase Keynote | Interrupt London

Channel: LangChain (@LangChain)

Link: Watch source video ↗

LangChain co-founder and CEO Harrison Chase argues that what a team really owns is not the model but the specialization layered on top: the harness, the context, and the way the system learns from use. Agents need to be open, compounding, and governed, and he points to monday.com, Harvey, and Orange as companies building the whole system around the model. He also walks through LangSmith, LangChain's tooling for watching and improving agents in production.

How Gen uses ChatGPT Work to turn data Into decisions

Channel: OpenAI (@OpenAI)

Links: Watch source video ↗ and Watch source video ↗ (two uploads of the same story)

Gen, the company behind Norton and Avast, uses ChatGPT Work to turn business data into decisions. Executive Howie Xu says that before ChatGPT Work they were walking and now they are running: Gen can understand how a new product's onboarding experience is working within two or three hours of release. The payoff is a shorter loop between something happening in the world and someone being able to act on it, instead of repeated handoffs across teams, functions, and time zones.

There are many futures. Which do you want to build?

Channel: OpenAI (@OpenAI)

Link: Watch source video ↗

OpenAI's short film asks a bigger question than any single product announcement: with tools getting cheaper and more capable every week, which future do we want to build? It is a brand piece, but its timing lands: on a day full of cheaper models, offline models, and agents at work, the open question is what we point them at.

Running a 120B model offline on Surface RTX Spark Dev Box with Intelligent Terminal

Channel: Microsoft Developer (@MicrosoftDeveloper)

Link: Watch source video ↗

A 120-billion-parameter model runs entirely locally on a Surface Laptop Ultra, using its NVIDIA RTX Spark chip with 128 GB of unified memory and a feature called Intelligent Terminal. The demo shows the machine doing complex work like writing new agent skills, with all context handled on-device and nothing going to the cloud. That matters for privacy, for working offline or in secure environments, and for latency — a frontier-class model that never leaves your laptop.

Phonon-2: 164MB Accurate Speech Recognition Model: Run Locally

Channel: Fahd Mirza (@fahdmirza)

Link: Watch source video ↗

Fahd locally installs and tests Phonon-2, a speech recognition model of just 164 MB described as the most accurate open English speech model under 900 MB. Accurate transcription on your own machine, in a file smaller than a short video clip, with no audio leaving your device. It pairs with the offline laptop story to sketch the emerging shape of AI: the cloud for the hardest problems, small capable models on your own hardware for everything else.

Automation vs Physical AI

Channel: Microsoft Developer (@MicrosoftDeveloper)

Link: Watch source video ↗

Automation repeats a task; physical AI gets smarter every time it runs. That is the distinction from Michael MacKenzie, Microsoft's VP of Product Management for Azure IoT and Digital Operations. Automation does the same thing the same way forever, while physical AI in factories, warehouses, and machines learns from each run and improves. The test for what gets sold as AI is simple: does the system get better the more it runs, or does it just repeat?

Customize your agent from a Chat Window?

Channel: Microsoft Developer (@MicrosoftDeveloper)

Link: Watch source video ↗

Using the Microsoft Foundry model catalog, you can find the best model for a specific use case, swap only the model deployment, and run the same evaluation over both options, capturing real cost and latency from real runs. The video argues the biggest model is quietly the wrong one more often than you would think. Turning model selection from a vibes-based decision into a measured one is unglamorous advice that saves real money.

Copy any app, Claude does SEO, and the top 10 repos

Channel: TheNextNewThingAI (@TheNextNewThingAI)

Link: Watch source video ↗

A roundup of the builder culture right now: cloning an existing app as a starting point, handing marketing chores like SEO to Claude, and watching which open source repositories the community is rallying around. Less starting from a blank page, more remixing, automating the boring parts, and learning in public from what is trending.