😺 Demis says chat came too soon—AI rivals unite – The Neuron

Welcome to the forefront of conversational AI as we explore the fascinating world of AI chatbots in our dedicated blog series. Discover the latest advancements, applications, and strategies that propel the evolution of chatbot technology. From enhancing customer interactions to streamlining business processes, these articles delve into the innovative ways artificial intelligence is shaping the landscape of automated conversational agents. Whether you’re a business owner, developer, or simply intrigued by the future of interactive technology, join us on this journey to unravel the transformative power and endless possibilities of AI chatbots.
Grant Harvey
April 12, 2026
 Sign Up · Advertise
Welcome, humans.
LM Studio just bought Locally AI, the app that lets you run open-source models straight on your iPhone, iPad, and Mac—offline, privately, and without the usual “please hand over your email, soul, and firstborn” setup. As part of the deal, Locally AI creator Adrien Grondin is joining LM Studio to lead native AI experiences across devices, which is a pretty clear signal that LM Studio wants to be more than your desktop’s local-model clubhouse.
Why it matters: local AI is creeping off the laptop and into your actual life. LM Studio said the goal is to make your own models and agents work “across your devices, wherever you go,” while Locally AI already supports offline, on-device use for models like Llama, Gemma, Qwen, and DeepSeek on Apple hardware. Translation: the future of AI might not just be bigger cloud models—it might also be your personal model, running on your devices, with your data staying put.
The bigger shift here is that private AI is getting consumer-friendly. What used to feel like a hobby for power users is starting to look like a real product category: cross-device, always available, and less dependent on Big Tech’s rented brain in the sky. If LM Studio nails the handoff between desktop and mobile, local AI could go from “cool demo” to “why am I paying monthly to summarize my grocery list?” 
Here’s what happened in AI today:
Meta’s AI app jumped to No. 5 after Muse Spark launched.
Andy Jassy doubled down on Amazon’s $200B AI spending plan.
Anthropic’s Mythos rollout sparked questions about whether safety or self-interest drove the limits.
OpenAI unveiled a child safety blueprint to combat AI-generated abuse.
Anthropic’s new agent software triggered a software stock sell-off.
P.S: Want to reach 675,000 AI-hungry readers? Click here to advertise with us.
P.P.S: Love robots? We’re starting a new robotics newsletter! Sign up early here.
Demis Hassabis just admitted something pretty striking: the AI boom didn’t unfold the way he wanted. Not because chatbots are fake, or useless, or overhyped. But because if he’d had his way, AI would’ve stayed in the lab longer solving science, medicine, and energy problems before becoming everybody’s browser tab.
In a new interview with Cleo Abram, Hassabis, the CEO of Google DeepMind, said the “best use case of AI” was always improving human health and accelerating scientific discovery. In his ideal version of history, AGI would’ve been built in a slower, more rigorous, more collaborative “CERN-like” way. (CERN is the wicked huge 16.5-mile particle accelerator in Switzerland.)
Instead, the opposite happened.
Language turned out easier than many researchers expected. Hassabis said even optimists thought it might take a few more breakthroughs to crack abstraction and conversation, but transformers got there faster than expected. Then OpenAI shipped ChatGPT, it went viral (even to OpenAI’s surprise), and the whole field got pulled into what he called a “ferocious commercial pressure race,” with geopolitics piled on top. 
You can hear the alternate timeline rattling around in his head. In that version, labs spend more time building many AlphaFolds, specialized systems that tackle huge scientific bottlenecks, while AGI advances carefully in the background. That, he said, would’ve been the cleaner rollout: fewer viral chatbot moments, more shots at curing disease or discovering new materials.
And to be fair, he’s not pretending the chatbot era was all downside. He gives it real credit:
Progress moved much faster
The public got hands-on access to frontier AI
Millions of users stress-tested these systems in the wild
Still, when Hassabis talks about AI with genuine awe, he’s not really talking about a better meeting summary. He’s talking about systems like AlphaFold, which he says is now used by more than 3 million scientists and may touch almost every future drug in some way.
Why This Matters: Consumer AI may be the loudest part of this wave, but Hassabis still thinks science is the main event. The chatbots got out first. The deeper ambition never changed.
So while the rest of the industry fights over whose assistant can book a restaurant fastest, Hassabis is basically arguing for something bigger. The most important AI of the next decade may be the kind you barely see at all.
Our Take: While I agree Hassibis’ approach would have been the safer route, the surprise AI boom we got, means money has flooded into the space at a previously unimaginable space. In the long run, I’m optimistic this will bring that science to reality years or decades sooner than the other path. That time means some number of diseases cured and historical problems solved long before it would have the other way. 
FROM OUR PARTNERS
At Arm Everywhere, Arm marked a major milestone for the compute platform: its expansion into production silicon with the Arm AGI CPU. Joined on stage by leaders from Meta and OpenAI, Arm outlined what this move means for the future of AI infrastructure. Built for the rise of agentic AI, this next step brings Arm into data center silicon for the first time—signaling a more integrated approach to performance, efficiency and scale across the AI stack.
Watch on Demand
Stop over-explaining, and start giving your AI examples of what you want. This is a rule we live by here at The Neuron. A lot of people write their AI a mini novel and hope it somehow nails the tone, format, and vibe. Usually, it doesn’t. 
If you want a summary in a certain style, a LinkedIn post that sounds like you, or notes cleaned up a specific way, showing one strong example usually works better than a long list of instructions.
That’s because AI is great at spotting patterns. A good example can quietly teach tone, structure, length, and detail all at once, without you having to spell out every rule. So instead of describing the output for five lines, paste a version you already like and say, “Make this new one look like that one.”
It’s a small shift, but it can make AI much more useful. You’ll usually get closer to what you want, with less back-and-forth and a lot less generic mush. The best AI users often aren’t giving better instructions, they’re just giving the model a better pattern to follow.
Total AI beginner? Start here (goes with this video).
Have a specific skill you want to learn? Request it here. 
New episodes air every week on: Spotify | Apple Podcasts | YouTube 

*Asterisk = from our partners (only the first one!). Advertise to 675K+ readers here!
OpenRouter Model Fusion sends your prompt to multiple models at once, then fuses the best parts of every response into one answer that outperformed each model’s own output in testing. No pricing details.
MemPalace gives your AI persistent memory across sessions by organizing conversations into navigable wings, halls, and rooms — built by Milla Jovovich (HN discussion), it scored highest on the LongMemEval benchmark and runs entirely locally. Free.
Hippo gives your AI agent memory that works like a human brain — important things stick, noise fades, and a “sleep” command compresses episodes into patterns (HN discussion). Free.
Finalrun tests your mobile app by reading the screen like a human — write a plain-English test in YAML and it taps, swipes, and types its way through your app on a real emulator, then hands you a pass/fail report with video. Free.
Ownscribe transcribes and summarizes your Zoom, Teams, or Meet calls locally so you can ask things like “what did Anna say about the deadline?” without anything leaving your machine. Free.
DocMason compiles your private decks, spreadsheets, and PDFs into a local knowledge base where an AI agent reasons over the evidence and points you to the exact file and page. Free.
Meta’s AI app climbed to No. 5 on the App Store after the company launched its new Muse Spark model.
Amazon CEO Andy Jassy defended the company’s aggressive AI spending, saying Amazon plans to invest about $200B this year, mostly in AI infrastructure and chips.
TechCrunch argued that Anthropic’s limited release of Claude Mythos raised questions about whether the company was protecting the internet or protecting itself.
OpenAI released a new Child Safety Blueprint aimed at addressing the rise of AI-generated child sexual abuse material.
Anthropic’s new agent software sparked a sell-off in software stocks as investors worried AI agents could replace parts of the SaaS stack. Again.

For years, the AI race has looked like three kids shoving each other off the jungle gym. Now they’re locking arms. 
According to a Bloomberg report, OpenAI, Anthropic, and Google are working together to detect and stop “adversarial distillation” attempts, especially from Chinese competitors.
The work is happening through the Frontier Model Forum, the industry group those companies launched in 2023. What started as a safety-focused coalition is starting to look more like AI’s neighborhood watch—except the neighborhood costs tens of billions to build.
The issue is distillation. In simple terms: you ask a powerful model enough smart questions, collect its answers, and use those outputs to train your own cheaper model. That can let competitors copy core capabilities without paying the original training bill—and potentially without inheriting the original safety guardrails.
Why now? Because the money is getting ridiculous. These labs aren’t protecting science projects anymore. They’re protecting giant businesses built on chips, cloud contracts, and subscriptions.
Anthropic is a great example. In February, the company said it had reached a $14B revenue run rate. Now, in its new Google / Broadcom compute announcement, it says that figure has jumped to $30B ARR. That makes this story feel less like abstract AI drama and more like old-fashioned corporate defense: when the revenue gets that big, everyone suddenly cares a lot more about model theft.
Why this matters:
AI competition is shifting from “who has the smartest model?” to “who can stop others from cloning it?”
Safety and geopolitics are merging. These companies aren’t just warning about lost revenue—they’re also arguing that copied models could spread powerful capabilities without the original safeguards.
The punchline: AI’s biggest labs still want to beat each other. They’d just prefer to do it the old-fashioned way—by spending absurd sums themselves, not by letting rivals speedrun the process with copied outputs.
That’s all for now.

Login or Subscribe to participate in polls.
P.S: Before you go… have you subscribed to our YouTube Channel? If not, can you?
Click the image to subscribe!
P.P.S: Love the newsletter, but only want to get it once per week? Don’t unsubscribe—update your preferences here.
The Neuron
Don't fall behind on AI. Get the AI trends and tools you need to know. Join 600,000+ professionals from top companies like Microsoft, Apple, Salesforce and more. 👇
Home
Posts
Authors
Home
Home
Advertise
Advertise
© 2026 TechnologyAdvice, LLC.
Privacy policy
Terms of use

source