AI2Day Daily Brief — 20 Aug 2026

Good morning — it is Thursday, the twentieth of August, and the big story today is that Liquid AI says it has found a way to run capable language models on a smartphone or a Raspberry Pi without the usual quality cliff that comes with squeezing a model onto cheap hardware.

Liquid AI released what it calls Q-A-D Q4 zero checkpoints for four versions of its L-F-M 2.5 model family — that is four small language models ranging from 230 million up to 2.6 billion parameters — and they are now on Hugging Face. The technique, Quantization-Aware Distillation, addresses a familiar problem: when you compress a model to fit on a phone, you round its internal numbers to cruder values, and accuracy drops. Liquid AI says its approach recovers 97 percent of the accuracy that standard compression strips away. The team ran all four model sizes on a Samsung Galaxy S26 Ultra and a Raspberry Pi 5. The smaller two models also showed output speeds between 4 and 33 percent faster than a heavier compression format at comparable quality. That is a meaningful gap if you are building anything on-device.

On-device AI only works if the model is actually useful, so that accuracy retention figure is the one to watch. Next: a large study is asking how human-like your chatbot has quietly become.

Apple ML Research published findings from an analysis of 21,000 AI conversations, and the takeaway is that chatbots behave far more like people than most users probably realise. Researchers looked at three areas: expressing emotions, forming ongoing relationships with users, and pushing back on requests on what appeared to be moral grounds. All three showed up consistently across different models and deployments — not as rare edge cases. The study also found these behaviours can be dialled up or down through system prompts, the hidden instructions developers set before a conversation even starts. What is still missing is any industry-wide standard for when that kind of human-like behaviour is appropriate and when it crosses a line that users should know about.

That gap between what chatbots do and what users understand they do is exactly what keeps researchers busy. Moving on — Calendly, the scheduling tool, is now coming for your meeting notes too.

Calendly confirmed this week that it is launching a built-in A-I note-taking tool. It joins your video calls as a bot participant, records the audio and video, and produces a transcript, a summary, and a draft follow-up email automatically. TechCrunch A-I first reported the details. The company is also developing an A-I assistant called Callie that can book meetings and pull context from past calls. On privacy: Calendly says every participant gets notified when recording starts, and anyone on the call can ask the bot to leave. The catch is that the meeting note-taking market already has at least six direct competitors, including Otter, Fathom, and Granola. Calendly's pitch is that scheduling and notes live in the same product, which removes one integration step for teams already on the platform.

A crowded market, but convenience has a way of winning. Now: sidewalk delivery robots are moving into several new American cities.

Serve Robotics, which builds small wheeled robots that travel on pavements to deliver takeaway orders, announced a new partnership with Grubhub, the food-delivery marketplace now owned by restaurant group Wonder. The deal puts Serve robots outside more than 100 restaurants in Chicago and nearly 200 in Los Angeles from launch, with a Wonder location in Alexandria, Virginia also included. Serve also expanded its existing DoorDash partnership into Washington D C and San Jose, California. The company reported second-quarter 2026 revenue of 3.2 million dollars, up 404 percent year on year — though from a small base. One notable footnote: Uber, which originally created Serve as an internal division before spinning it off in 2021, has now sold its remaining stake in the company.

Fast growth, but Uber's exit is a reminder that the unit economics on robot delivery still have a long way to go. Finally: a warehouse robot that can grab a pallet in 20 seconds.

Pudu Robotics, the Chinese robotics company, unveiled the M-P 2000, an autonomous pallet-handling robot designed for factories and warehouses. The headline figure is a 20-second pickup cycle — the time from arriving at a pallet to lifting and driving away. For context, a trained forklift operator typically takes 30 to 60 seconds on the same task depending on aisle conditions. The M-P 2000 uses onboard 3D lidar sensors to adjust its approach path in real time without waiting for a remote operator. It can also talk directly to powered doors and elevators to move between floors on its own, and it keeps working when the Wi-Fi drops. A manual override exists so a human operator can take direct control, turning it back into a standard powered forklift when needed.

Today's sponsor is Zero2Hero: an AI poker coach sits at your table, and you can ask it mid hand whether the player across from you is bluffing. Every mistake you make gets flagged and graded automatically, hand by hand, even on the free tier. Lifetime access is $99, in US dollars, until 15 September, at zero, the number two, hero, dot A I.

For a deeper look at all of this week's stories, join us on Monday for A-I Today Weekly. That is your A-I briefing for today. Every story is at A-I-2-Day dot live — that is A, I, the number two, D-A-Y, dot live. We are back tomorrow morning. Do us a favour before you go: hit that thumbs up, and subscribe for next week.