š The Prompt Factory
Pope Leo š identifies AI as one of humanity's greatest challenges.
š° This Week's AI Highlights
This week's AI highlights include 19pine.ai, a practical voice AI application that negotiates with customer service representatives on your behalf to handle refunds, cancellations, and fee disputes. The model landscape continues evolving with Mistral Medium emerging as Europe's competitive entry, Google's Gemini Pro 2.5 showing exceptional coding abilities, and OpenAI maintaining dominant market share. Creative tools are democratizing sophisticated content creation, with new systems for AI-generated animations, HeyGen's technology that transforms static images into expressive avatars, and fully AI-generated game worlds. For developers, GitHub project Mem0 (boasting 28.8k stars) offers efficient memory management for AI agents, while Stanford's CS25 course features lectures from leading AI researchers like Geoffrey Hinton. We also touch on deeper questions, including philosophical discussions about AI consciousness and welfare, alongside Pope Leo's identification of artificial intelligence as one of humanity's major challenges.
Full disclosure: this TLDR was written by AI.
š„° Fun things
A fun Gemini animation: This is cool and *very* sophisticated.
Link: Gemini
Our Take: Switch to the code view and see the number of lines of code that were written by Gemini. Itās pretty amazing that we have AI writing this kind of stuff.
AI Game Worlds: Welcome to a world in which the entirety of the content is generated by AI.
Link: Enigma Labs
Our Take: This really is the path towards the Star Trek āholodeckā, where we enter an AI-generated world. Itās basic gaming right now, but itās easy to see how this will rapidly improveā¦ and when we add VR/AR headsets into the mix we get somewhere interesting, or weird depending on your perspective.
š¤ Agents
Mem0 memory for your agent: Mem0 is an LLM-powered memory pipeline to ensure only the most relevant facts are stored, keeping tokens low and latency fast.
Link: Github
Our Take: 28.8k stars is not to be argued with - this looks pretty neat.
š¦¾ Agents that Get Stuff Done
Here we focus not on AI technology itself, but on the practical applications of it.
Imagine you had an AI assistant that actually enjoys arguing with customer service reps, so you donāt have to? Thatās exactly what 19pine.ai does: āMost disputes with businesses can be resolved with a simple conversation, but they often turn into a frustrating ordeal. Many people, worn down by long hold times, endless transfers, and unhelpful agents, eventually give up on fighting for whatās rightfully theirsā¦ Itās not just about the moneyāitās the feeling of being dismissed.ā That certainly resonates!
19pine.ai phones your adversary (!) and uses voice AI to negotiate and claim refunds, cancel subscriptions, waive fees and reduce bills. Itās a good example of how voice AI is going to transform so many things.
But how are organisations going to respond to an army of bots commissioned by consumers to do their bidding? If the caller is a bot, why should the agent not also be one? Let the bot wars commence, whilst we sit back and have a glass of something cold to drinkā¦
šØ UI
Animation prompt builder: Get an LLM to create animations for you with this handy prompt builder.
Link: Aurachat
Our Take: The animations on the barnacle.ai website were created using an LLM. Until recently doing this kind of stuff was the preserve of a very niche and specialist skill, but now anyone can do it (including me!). But prompting an LLM in natural language is a bit hit-and-miss ā you typically end up with random animations. Instead, this page (scroll to the bottom) includes a handy prompt-builder where you can choose what animation you want and it generates the natural language prompt for you to copy into ChatGPT (or whatever).
š Papers
Advances and Challenges in Foundation Agents: Everything you wanted to know about building agents.
Link: arXix
Our Take: Notice the title is āAdvances and CHALLENGES in Foundation Agentsā. Regardless, this is 264 pages, so quite some detail!
š Education
Stanford CS25: Learn from the best.
Link: Stanford
Our Take: CS25 features rockstar AI researchers as lecturers, including Geoffrey Hinton, Ashish Vaswani and Andrej Karpathy. Each week, it dives into the latest breakthroughs in AI, from large language models to applications in art, biology, and robotics. Itās also free!
š® Model Watch
Mistral Medium: Finally, a top-grade AI model from Europe.
Link: Mistral Blog
Our Take: Weāve been used to the USA and China making all the frontier class modelsā¦ now itās time for Europe to play a role. According to Artificial Analysis (the LLM testing outfit), āMistral is back amongst the leading non-reasoning models with Medium 3 rivalling Llama 4 Maverick, Gemini 2.0 Flash and Claude 3.7 Sonnetā. Itās an excellent model, competing with the very best, but Mistral also hint at an upcoming Mistral Large. So it looks like Europe is coming out punching here. About time! I have a soft spot for all things French and that Mistral named their version of ChatGPT āLe Chatā makes me smile!
Google Gemini Pro updated: Google updated their Gemini Pro 2.5 model and everyoneās gone ga-ga about its coding abilities.
Link: Gemini Blog
Our Take: Gemini Pro 2.5 was already really good for coding, but this latest update has made it amazingāitās reached the top place in all LMArena leaderboards. Whatās more, as a model classed as āexperimentalā, we can all use it for free!
NVidia Parakeet Speech-to-Text: This one went straight to the top of the speech leaderboard.
Link: Hugging Face
Our Take: Speech recognition is getting really good. Itās interesting that this is both open source and currently leads the ASR leaderboard.
HeyGen Avatar IV: Turns a single static image (e.g. a photo of you) into a moving avatar with lip-sync, emotion and head movement.
Link: HeyGen
Our Take: Iāve always hated videoing myself. Maybe now I donāt need to!
Model share: See a breakdown of the marketshare of each model.
Link: x
Our Take: No surprises that OpenAI is the biggestā¦ but would you have predicted by this amount?
o4-mini RL Fine-Tuning:
Link: OpenAI Docs
Our Take: Until recently fine-tuning was virtually always āsupervised fine-tuningā, i.e. example-based, where you provide examples of question/answer pairs. However, the community is getting excited about reinforcement learning, where instead of training on fixed ācorrectā answers you rely instead on a programmable reward algorithm that scores every candidate response. The idea that o4-mini can now be reward-trained is quite exciting to us geeks!
š§ AGI ā Are we there yet?
AI Welfare: Should we be concerned about the potential consciousness and experiences of AI models? Should we be concerned about model welfare?
Link: Anthropic
Our Take: I dunno, this is weird. But it raises interesting questionsādefinitely one for the philosophers! The blog article references the paper āTaking AI Welfare Seriouslyā that goes into much more detail.
People and possibilities in the age of AI: A very comprehensive review of the challenges and opportunities that AI presents.
Link: UN
Our Take: From the report: āWe are at a crossroads: while AI promises to redefine our future, it also risks deepening the divides of a world already off balance. Are we on the verge of an AI-powered renaissanceāor sleepwalking into a future ruled by inequality and eroded freedoms?ā Big questions and a lot of very thoughtful content.
Pope Leo & AI: The new pope identifies AI as one of the main challenges for humanity in his first utterances.
Link: The Independent
Our Take: āIn our own day, the church offers everyone the treasury of its social teaching in response to another industrial revolution and to developments in the field of artificial intelligence that pose new challenges for the defence of human dignity, justice and labour,ā Leo said.
The Prompt Factory is a newsletter from Barnacle Labs, sharing AI insights and discoveries our team finds fascinating.