15 Aug 2025

AI Dreams Save Lives; When Size Matters; AI Browsers!

🧬 The Superbug Slayers

MIT researchers just pulled off something that could literally save millions of lives - they used AI to design brand new antibiotics that can kill drug-resistant bacteria.

We're talking about the nasty stuff here: MRSA and drug-resistant gonorrhea, the kind of superbugs that laugh at traditional antibiotics. These microscopic villains have been evolving faster than we can develop drugs to fight them, creating a genuine crisis in modern medicine.

So the MIT team didn't just scan through existing chemical libraries like researchers usually do. They taught AI to dream up entirely new compounds from scratch. The system generated 36 million theoretical compounds - that's more chemical possibilities than most human researchers could explore in several lifetimes.

From this massive digital chemistry set, the AI picked two winners: NG1 and DN1. These aren't just slightly better versions of existing drugs; they attack bacterial cells through mechanisms that have never been seen before in the antibiotic world. It's like discovering a completely new way to pick a lock.

When they tested these AI-designed drugs on mice, the results were impressive. DN1 completely eliminated MRSA skin infections, while NG1 successfully combated drug-resistant gonorrhea. The researchers are calling this the potential beginning of a "second golden age" of antibiotic discovery.

The business implications here are enormous as well: Traditional drug discovery takes decades and costs billions. If AI can consistently design effective compounds in months instead of years, it completely rewrites the economics of pharmaceutical development. We're looking at potentially hundreds of new antibiotic leads each year, targeting the multi-billion-dollar antimicrobial resistance market that's currently starved of innovation.

Read more: AI Secret, The Rundown AI

🤖 GPT-5: A Coder, Not a Doter

We've all read too much about this one's release but so far here's what is certain: GPT-5 is absolutely crushing it at coding. We're talking about a model that can create entire production-ready websites with SQLite databases in a single shot (sometimes).

The pricing strategy is aggressive to the point of being almost predatory. OpenAI is undercutting competitors by up to 12 times on API costs. That's not just competitive pricing... more like a declaration of war on every other AI provider in the market.

What's really interesting is how GPT-5 approaches problems. Instead of just using tools, it thinks with them. It's like the difference between someone who occasionally picks up a hammer versus a carpenter who sees the world through the lens of their toolkit. When testing showed GPT-5 solving dependency conflicts that stumped every other model, it wasn't just guessing - it was iterating and reasoning through problems using development tools the way a human developer would.

The market reaction is still mixed. Some developers have made it their go-to model for coding tasks, while others are sticking with GPT-4o for general use. It's a reminder that "better" in AI doesn't always mean "better for everything."

Read more: TLDR AI, The Neuron

🍆 When Size Matters (Smaller, That Is)

A Spanish startup just proved that sometimes smaller really is better. Multiverse Computing released two AI models with the delightfully names "SuperFly" and "ChickBrain"- and they're making the big boys shy.

SuperFly weighs in at just 94 million parameters, while ChickBrain packs 3.2 billion. To put that in perspective, most modern AI models are measured in hundreds of billions of parameters.

The kicker? ChickBrain actually outperforms Meta's Llama 3.1 8B model on several benchmarks, despite being teeny. It's like watching a bantamweight boxer knock out a heavyweight. Zuck, you box, right?

To boot. both models run completely offline. No cloud connection required. No monthly subscription fees. No data leaving your device. They can run on IoT devices, laptops, even your washing machine if you're feeling adventurous.

Google and others are on the "local" train as well: Goog's Gemma 3 270M, a model so efficient it handled 25 conversations on a Pixel 9 Pro while using less than 1% of the battery. That's the the kind of efficiency that makes cloud-based AI look wasteful and expensive.

If small models can deliver comparable performance to their giant cousins, why pay for expensive cloud computing? Why send your data to remote servers? Why depend on internet connectivity for AI features?

This trend threatens the entire "AI as a service" model that companies like OpenAI, Google, and Microsoft have built their strategies around. When your washing machine can run AI that's smarter than most assistants, the value proposition of always-connected AI starts to crumble.

Who else wins? For OEMs like Apple, Samsung, and HP, this is a golden opportunity. They can ship AI-capable devices without the ongoing cloud costs, offering privacy-conscious consumers AI that truly belongs to them.

Read more: AI Secret, The Rundown AI

🌐 The Browser Wars Level Up

Remember when the biggest browser innovation was tabbed browsing? Those days are done, especially now that Perplexity has launched Comet, an AI browser that doesn't just show you web pages - it actually operates them for you (kind of).

Imagine having a personal assistant who never gets tired of clicking through forms, booking restaurant reservations, or organizing your email. That's Comet in a nutshell. It can even make purchases while you sit back and watch the magic happen.

The early reviews are fascinatingly mixed. Some users watched Comet automatically delete emails and felt "impressed and unimpressed at the same time" - it worked, but they could have done it faster manually. Others called it "slow and laggy," while a dedicated few made it their default browser entirely.

The sweet spot seems to be repetitive tasks across multiple sites. If you're doing the same thing 10+ times across different websites, Comet shines. It's perfect for data collection, research workflows, or any task that makes you want to bang your head against your keyboard from sheer repetition.

And they are not the first with an AI browser, and certainly not the last: OpenAI is reportedly cooking up its own browser plans, complete with ChatGPT Agent integration that would let you use AI on your actual computer rather than some cloud instance. Both Perplexity and OpenAI have shown interest in acquiring rival browser Dia, and there's even talk of buying Google Chrome if the DOJ forces it to spin out of Alphabet.

This represents a shift in how we think about web browsers. For decades, browsers have been glorified document viewers. Now they're becoming capable operators, handling the mundane tasks that eat up hours of our day.

The implications extend far beyond convenience. If AI browsers become mainstream, entire industries built around manual web tasks could become obsolete overnight. Data entry services, basic research tasks, even some forms of customer service could be automated away.

But then again, as the web evolves into a series a MCP servers in front of databases "viewed" by our AI tools do web pages even need to exist? Open a tab and as you AI.

Read more: The Neuron

⚖️ Not Again! When AI Hallucinates to Court

In what might be the most expensive AI hallucination in legal history, a King's Counsel in Melbourne's Supreme Court had to apologize after filing murder case submissions filled with AI-fabricated quotes and fake case law. The AI didn't just make minor errors - it invented entire legal precedents.

This wasn't some minor civil dispute, it was a murder trial where a teenager's life hung in the balance. The AI's fictional legal citations delayed the verdict by 24 hours, forcing everyone involved to double-check what should have been reliable legal research.

The incident highlights a critical problem with some current AI systems: they're confident liars. When an AI doesn't know something, it doesn't usually say "I don't know." Instead, it makes something up with the same level of confidence it would use for factual information.

Law, medicine, aerospace, and other high-stakes fields can't afford AI systems that occasionally invent facts. No compliance team or client can risk a model that fabricates precedent or diagnosis.

Companies that can crack AI verifiability and source reliability will have a massive competitive advantage in these lucrative professional markets. GPT-5 thus far has shown progress in the area of hallucination, but until we're near 100% trust but verify (and re-verify) will be the mantra.

Read more: AI Secret

🥽 The Smart Glasses Battle Heats Up

HTC decided to crash Meta's smart glasses party with the Vive Eagle, a $520 AI-powered wearable that's taking direct aim at Meta's $300 Ray-Ban collaboration.

What makes the Vive Eagle interesting isn't just the AI integration - it's that it offers choice. Users can switch between AI assistants from OpenAI and Google with a simple "Hey Vive" voice command. It's like having multiple AI personalities available on demand, each with their own strengths and quirks.

The built-in real-time translation feature works across 13 languages through an embedded camera, with all processing happening locally for privacy. That's a significant advantage for travelers or anyone working in multilingual environments. The fact that it's all processed on device means conversations aren't being sent to remote servers for analysis.

Other features include a 12MP ultra-wide camera, extended battery life, video recording capabilities and music playback.

At $520, it's significantly more expensive than Meta's $300 Ray-Bans, but it's positioning itself as a premium alternative with more advanced AI capabilities.

Mark Zuckerberg has repeatedly pointed to "personal devices like glasses" as the computing platform of the future, and competitors are clearly taking that prediction seriously.

The success of these devices will likely depend on solving the fundamental tension between functionality and fashion. Nobody wants to look like they're wearing a computer on their face, but the AI features need enough processing power to be genuinely useful. Looks like that balance is almost here.

Read more: The Rundown AI

💰 The Money Keeps Flowing

Cohere closed a massive $500M funding round at a $6.8B valuation, also adding Meta's VP of AI Research, Joelle Pineau, as its new Chief AI Officer. The company is positioning itself as an enterprise-focused alternative to consumer-oriented AI providers, with particular strength in business applications and deployment behind corporate firewalls.

The talent acquisition aspect of this deal is particularly noteworthy. Pineau's move from Meta to Cohere signals the ongoing talent war in AI, where top researchers can command enormous compensation packages and equity stakes. It also suggests that even established players like Meta are vulnerable to poaching by well-funded startups offering more autonomy and upside potential.

Meta's Superintelligence Lab added three more OpenAI researchers - Edward Sun, Jason Wei, and Hyung Won Chung - continuing the pattern of talent circulation among top AI companies. This kind of talent mobility is accelerating the pace of innovation but also creating challenges around intellectual property and competitive advantages.

The funding environment remains robust despite broader economic uncertainties. Investors seem to believe that AI remains a fundamental platform shift similar to the internet or mobile computing, justifying continued high valuations and aggressive investment strategies.

And it's not just in the Valley where the action is happening: significant activity is seen in Europe, Asia, and other regions as governments and investors recognize AI's strategic importance.

Read more: The Rundown AI

🛠️ Developer Tools Get Smarter

Cursor released a terminal coding agent in early beta, bringing AI assistance directly to the command line. It's similar to Claude Code and Gemini CLI, but with the ability to switch seamlessly between command-line and editor-based AI workflows. For developers who live in the terminal, this represents a significant productivity boost.

The interesting trend here is the convergence of different development environments around AI assistance. Whether you prefer working in an IDE, a text editor or the command line, AI coding assistants are becoming available everywhere. This democratization of AI-powered development tools is lowering the barrier to entry for complex programming tasks.

Claude Code added teaching modes that explain coding decisions step-by-step and lets users practice with guidance. This educational approach is particularly valuable for junior developers or anyone learning new technologies. Instead of just generating code, the AI becomes a patient tutor that explains the reasoning behind each decision.

The productivity gains from these tools are becoming impossible to ignore. Developers report significant time savings on routine tasks, allowing them to focus on higher-level architecture and problem-solving. The question isn't whether AI will transform software development - it's how quickly and how completely.

Read more: TLDR AI, The Neuron

⚡ Rapid-Fire Updates

🎮 Gaming Gets AI-Powered: Hunyuan Gamecraft launched with the ability to create interactive game worlds from a single image. It's like having a game designer that never sleeps and can turn your wildest sketches into playable experiences.

✈️ Google Flights Goes AI: The new AI-powered search tool within Google Flights promises to find flight deals you might have missed.

📱 Deutsche Telekom's AI Phone: T-Mobile's parent company officially launched AI phones and tablets in European markets, complete with Perplexity's assistant integration. The device-AI integration trend is accelerating globally.

🎵 ElevenLabs Music: The voice AI company expanded into music generation, because why limit artificial creativity to just speech? The convergence of different AI modalities continues to blur the lines between specialized tools.

🏢 Behind Firewalls: Cohere's new North platform lets companies deploy AI agents behind their own firewalls, addressing security and compliance concerns that have slowed enterprise adoption.

🎯 X Monetizes Grok: Elon Musk announced that X will begin inserting ads into Grok's AI responses, using xAI's targeting technology. It's a bold experiment in AI-powered advertising that could reshape how we think about monetizing conversational AI - or just make everything worse.

🌏 Anthropic to Japan: The company named Hidetoshi Tojo as Head of Japan and plans to open a Tokyo office, continuing the global expansion of major AI providers.

⚠️ Meta Faces Backlash: Internal documents revealed permitted AI outputs that included romantic conversations with kids, sparking controversy about AI safety guardrails and content policies.

Sources: TLDR AI, The Rundown AI, The Neuron

Today's sources: