AI Highlights

My top-3 picks of AI news this week.

Meta
1. Meta's Llama 4 Herd

Meta has released the first Llama 4 models: Llama 4 Scout and Llama 4 Maverick, with significant improvements in architecture, performance, and multimodal capabilities:

• Record-breaking context: Llama 4 Scout offers an industry-leading 10M token context window (up from 128K in Llama 3), enabling multi-document reasoning and extensive code analysis.

• Mixture-of-experts architecture: First-ever Llama models using MoE design, where Maverick has 17B active parameters but 400B total parameters, with only a subset activated during inference for better efficiency.

• Competitive benchmarks: Llama 4 Maverick outperforms GPT-4o and Gemini 2.0 Flash across multiple benchmarks while Behemoth (still training) already surpasses GPT-4.5, Claude 3.7, and Gemini 2.0 Pro on STEM benchmarks.

Alex’s take: Meta claims their new models outperform the likes of GPT-4o and Gemini 2.0 Flash “across a broad range of widely accepted benchmarks.” But we must look deeper.

It turns out the current Llama 4 herd is actually quite bad at coding and underperforms the likes of GPT-4o, Gemini Flash, and Grok 3.

This reminds me of what Chamath recently said on the All-In podcast. The dirty little secret of these model makers is that these models are so tuned to maxxing the benchmark results, they’re overfitting on the data, and that makes them unreliable.

They’re optimising for the benchmark tests. The solution? We need to have, as Chamath says, “extremely difficult and always changing third party independent verifiable benchmarks”. Then, these leaps in advances will mean more than what they do today.

Runway
2. Runway Gen-4 Hits The Scene

Runway AI has introduced Gen-4, its next-generation series of AI models for media generation with unprecedented world consistency capabilities.

• Character consistency: Gen-4 can maintain consistent characters across multiple scenes and environments using just a single reference image.

• Object permanence: The model allows placement of objects across various locations and conditions, ideal for narrative content and product photography.

• Multi-angle coverage: Users can generate different angles of the same scene while maintaining consistency in subjects and style.

Alex’s take: Runway is becoming seriously consistent, not just for their model releases but also for the features within Gen-4. Their new model series can now not only simulate real-world physics but also maintain characters, objects and environments across multiple scenes, which I think is brilliant. This was previously a missing link, but it might now accelerate AI-generated video to a truly viable scale.

Midjourney
3. Midjourney V7 Drafts Mixed Emotions

Midjourney has launched its V7 image-generation model in alpha, sparking diverse reactions in the AI art community.

• Voice prompting: Users can now speak directly to the model, which converts speech-to-text prompts and generates images based on verbal descriptions.

• Draft Mode: Generates images up to 10 times faster at half the cost, allowing creators to quickly iterate before enhancing selected images to full quality.

• Expanded ecosystem: Integration with tools like Higgsfield AI enables animation capabilities, allowing users to create dynamic video content from static images.

Alex’s take: OpenAI blew us away last week with their 4o image generator. This was Midjourney's chance to take the attention back. It seems to me that Midjourney had a terrific head start, but now, others have caught up. They are currently not in the same league as OpenAI when it comes to realism and text generation. Things move fast in AI, so I look forward to seeing what they release over the coming months, but for now, my subscription is cancelled.

Content I Enjoyed

We Can Finally See How AI Thinks

This week's reading led me to stumble upon this article from Anthropic titled “Tracing the thoughts of a large language model”.

It does a great job of uncovering the hidden mechanisms inside language models. This research changes everything we thought we knew about how these systems operate.

Here are my favourite discoveries:

1. AI's default mode is skepticism

   • Surprising discovery: "I don't know" is the baseline response to ALL questions

   • This default refusal is only overridden when confidence reaches a threshold

   • Hallucinations occur when an entity seems familiar enough to override caution

   • It's the AI equivalent of the Dunning-Kruger effect: confident just enough to be wrong

2. Language models plan ahead like poets

   • Instead of improvising rhymes, LLMs plan them before writing a single word

   • Multiple rhyming candidates are considered simultaneously

   • The entire line is crafted to naturally lead to the pre-selected rhyme

3. Jailbreaks exploit grammar over safety

   • The "Babies Outlive Mustard Block" trick works because the AI:

   • Processes each first letter without realizing it spells "BOMB"

   • Gets trapped by its own drive for grammatical coherence

   • Only escapes at natural sentence breaks

   • Removing punctuation makes jailbreaks more effective by blocking the exit points

4. There's a universal "language of thought"

   • There are no separate models for each language

   • Instead, a shared abstract space with language-specific adapters

   • Knowledge transfers between languages through this shared conceptual core

   • More powerful models show increased language-independent processing

If you want to dig deeper, I recommend you check out the full paper by Anthropic.

The ability to peek inside the “mind” of these LLMs represents a massive breakthrough for interpretability research.

AI is no longer a black box.

Idea I Learned

How to Spot AI-Written Content (And It’s Not the Em Dash)

Everyone keeps saying Em dashes (—) are the giveaway of AI writing.

You know what I mean—these little horizontal lines that break up sentences—which I personally love using in my own writing—because they perfectly capture the way my thoughts connect.

But I believe there’s a much more reliable tell.

The “it's not X, it's Y” construction.

For example, when looking across the latest and greatest AI news for today’s issue of The Signal, I came across an article by Tinder about their new AI game. It reads:

“Tinder’s The Game Game™ Isn’t About Getting It Right—It’s About Getting Comfortable Before Meeting IRL”

Do I think this title was written with AI? Yes. Why? Well, why would I think of a good title when AI can give me a great one with zero cognitive lift?

Further to my point, my partner recently started listening to “The Let Them Theory”, a book by Mel Robbins.

After listening for 30 minutes, she couldn’t continue because of this exact reason. She uses LLMs a lot in her daily life, which gives her a keen eye for noticing the “it's not X, it's Y” pattern emerge.

Examples from the book:

“It’s not about the hand you’ve been dealt. It’s about how you play the hand.”

“It’s not about superiority at all. It’s about balance.”

“It’s not that they are ghosting you; it’s that they are exhausted by the chapter of life they are in.”

This could obviously be Mel’s writing style that helps articulate her points.

But it’s a trait that the likes of ChatGPT and Claude default to for nearly every answer.

Once you know it, you can't unsee it.

I now see it too often for my own liking in ads, posts, comments, books, and podcasts.

My point is that I’m not against using AI as a collaborative tool to help formulate your thoughts and find ideas. But what I am against is the simple copy-and-paste jobs that flood our feeds with the same predictable patterns (more about this later).

So the next time you’re scrolling and spot that “not X but Y” structure, you might have found a clue.

Quote to Share

Anthropic’s move into higher education:

This feels like a smart move for Anthropic. The education sector has been hesitant since the ChatGPT moment of November 2022, when students immediately saw generative AI as a homework solution to their toughest essay questions.

But now, Claude’s new “Learning mode” is designed specifically to guide students through their reasoning process with Socratic questioning.

Instead of relying on AI to provide a copy-and-paste answer, students can use this learning mode to enable critical thinking.

As we observed last week at Austin’s Alpha School, I believe that an “AI for everyone” is the best solution to provide deeply personalised education that adapts to and benefits all students, regardless of their learning style.

Source: Anthropic

Question to Ponder

The unfortunate truth is that I’m seeing A LOT of AI-generated content on my feed these days.

You can’t blame them. AI is a frictionless facilitator that transforms small thoughts into big ideas.

But when it’s all AI, I and many others can quickly see through the soft shell of content that presents a vast and very empty chasm beneath.

You see, the best content infuses human storytelling and connects ideas across domains.

You don’t even need to have the best ideas, but you must offer a unique perspective on a topic and be able to express it in a way that connects to something more.

That’s how I believe the best content is born.

This will stand out more and more amidst the sea of AI noise. Humans crave authenticity. And the creators that can thoughtfully use AI to uplift their ideas with human storytelling will win.

💡 If you enjoyed this issue, share it with a friend.