Sebastian Raschka: From DeepSeek-V3 to Kimi K2: A Look At Modern LLM Architecture Design

This article examines the structural changes and architectural developments in modern Large Language Models (LLMs), such as DeepSeek-V3, OLMo 2, Gemma 3, and Llama 4, rather than focusing on benchmark performance or training algorithms. The author details key innovations, including Multi-Head Latent Attention (MLA), Mixture-of-Experts (MoE), various normalization layer placements (Pre-Norm, Post-Norm, and QK-Norm), and sliding window attention, which primarily aim to enhance computational efficiency, memory usage, and training stability.

Paul Levchuk: The Metric Tree Trap

The article defines a Metric Tree as a hierarchical decomposition of a top-level business goal into measurable drivers, acknowledging its value primarily for visualisation and team alignment of key performance indicators. However, the author critically argues that Metric Trees are unreliable for making robust decisions, as they frequently obscure crucial operational insights due to issues such as contradictory metric definitions, inconsistent granularity, hidden trade-offs, and confounding factors, making the effective identification of key drivers, root cause analysis, and accurate prioritization challenging. To mitigate these “traps” and ensure reliable conclusions, the author strongly advises pairing Metric Tree insights with rigorous root cause analysis, scenario testing, and a thorough cost-benefit assessment.

What makes an AI startup succeed isn’t the tech—it’s the team. It turns out, the most successful companies aren’t always the ones with the best product. Instead, they’re the ones with the best teams. Teams who trust each other and can adapt rapidly. Dave Blundin (Managing Partner of AI venture fund, Link-XPV) quotes legendary VC Fred Wilson: “If they pass the test of being best friends and technical co-founders, I invest—even if the idea is stupid—because their idea will change, but the people won’t.”

1. The Number One Failure Mode

When founders aren’t deeply aligned, the first pivot often kills the company. That’s the number one failure mode for startups: somebody bails because they can’t handle the uncertainty. But they’ll succeed in the end if they commit and stick with it, together.

I remember the early days of Singularity University’s Graduate Summer Studies Program (GSP). We put a hundred alpha males and alpha females into a room and said, “Find other co-founders amongst yourselves and start a company based on exponential tech.” This was very different from Y Combinator (started roughly at the same time). In Y Combinator, the teams entered with preexisting relationships. In retrospect, the failure mode at SU’s incubator was predictable. There was no glue, no shared passion as these teams came together over the course of a few weeks, versus a few years.

In fact, the companies that did succeed in the GSP were the ones where founders became friends and stuck together through every pivot.

2. You Can’t Buy Founder Chemistry

People ask, “Why should they be best friends? Why should they have a relationship spanning years?”

Here’s why: Meta, Google, and OpenAI are raiding companies and stripping out talent. If you’ve started a company with some stranger and it’s been six months, when someone offers you a huge signing bonus, you’re going to leave.

But if you’ve started a company with your best friends—people with whom you have real history—you’re not going to abandon them. That’s critical.

Look at Zuck’s AI talent poaching spree. Meta tried to buy Safe Super Intelligence (SSI) for billions but got rebuffed. Now they’re throwing $100 million signing bonuses at executives like Daniel Gross and Nat Friedman. In the AI talent wars, you can throw money at researchers, but you can’t recreate the magic of bonded co-founders building late into the night.

3. Ideas Pivot, But Relationships Last

Fred Wilson, the legendary VC from Union Square Ventures, has a simple philosophy: he backs teams of best friends who are technical co-founders. His logic is: even if the idea is terrible, he’ll invest because great teams pivot fast, but you can’t change relationships overnight.

Take Israel’s startup ecosystem: five times higher success rate per capita than anywhere else. Why? The military service that predates most of the startups has created lifelong bonds. Marching through the desert, suffering together, appreciating college more, and then starting companies while still in school. They’re older but more bonded.

4. A Test for Strong Relationships

Even if you haven’t gone to school with your co-founders, there are ways to test compatibility.

Here’s mine: People you’d sit next to on a 12-hour flight in coach. How would you feel when you get off that plane? Are you exhausted or energized? That’s how it’ll feel doing a startup together. If you’re with the right people—those you click with—you’ll come off energized after talking about everything under the sun for 10-12 hours.

The future belongs to those bonded teams who can pivot together, stick together, and build together.

The dots were always everywhere, now they can finally connect them.

I’ve gotten a fair number of questions about the Planatir news, so let me lay out a few key things to keep in mind.

First:

The U.S. government has had this kind of capability—the ability to know anything about you, almost instantly—for decades.

Yeah. Decades.

Ever wonder how, three seconds before a terrorist attack, we know nothing, but three seconds after, we suddenly know their full bio, travel record, high school GPA, what they had for breakfast, the lap dance they got the night before, and the last time they took a dump?

Yeah. Data collection isn’t the problem. It never has been. The problem is, and always has been, connecting the dots.

The U.S. government vacuums up data 24/7. Some of it legally. Some of it… less so. And under the Trump Regime, let’s be honest—we’re not exactly seeing a culture of legal compliance over at DHS, the FBI, or anywhere else. Unless Pete Hegseth adds a hooker or a media executive to a Signal thread and it leaks, we’re not going to know what they’re doing.

But the safest bet? Assume Title 50 is out the f*ing window.

For the uninitiated: Title 50 governs U.S. intelligence operations. It prohibits turning that machinery—especially things like NTM (national technical means) or the NSA’s bulk intercept tools—against Americans.

There are exceptions. But they’re narrow. And they exist precisely to prevent a situation where someone like TACO can ruin your life because he’s having a hissy fit and is enabled by a ketamine-fueled choade with root access.

And yet—here we are.

In an era where the rhetoric of innovation is indistinguishable from statecraft, Code Dependent does not so much warn as it excavates. Madhumita Murgia has not written a treatise. She has offered evidence, damning, intimate, unignorable. Her subject is not artificial intelligence, but the human labor that props up its illusion: not the circuits, but the sweat.

Reading her work is like entering a collapsed mine: you feel the pressure, the depth, the lives sealed inside. She follows the human residue left on AI’s foundations, from the boardrooms of California where euphemism is strategy, to the informal settlements of Nairobi and the fractured tenements of Sofia. What emerges is not novelty, but repetition: another economy running on extraction, another generation gaslit into thinking the algorithm is neutral. AI, she suggests, is simply capitalism’s latest disguise. And its real architects, the data annotators, the moderators, the ‘human-in-the-loop’, remain beneath the surface, unthanked and profoundly necessary.

The subtitle might well have been The Human Infrastructure of Intelligence. The first revelation is that there is no such thing as a purely artificial intelligence. The systems we naively describe as autonomous are, in fact, propped up by an army of precarious, low-wage workers, annotators, moderators, cleaners of the digital gutters. Hiba in Bulgaria. Ian in Kibera. Ala, the beekeeper turned dataset technician. Their hands touch the data that touches our lives. They are not standing at the edge of technological history; they are kneeling beneath it, holding it up. Many of these annotators are casually employed as gig workers by the US$ 15 billion valued Scale.AI.

Artificial intelligence is transforming the entire pipeline from college to the workforce: from tests and grades to job applications and entry-level work.

This is a hard time to be a young person looking for a job. The unemployment rate for recent college graduates has spiked to recession levels, while the overall jobless rate remains quite low. By some measures, the labor market for recent grads hasn’t been this relatively weak in many decades. What I’ve called the “new grad gap”—that is, the difference in unemployment between recent grads and the overall economy—just set a modern record.

In a recent article, I offered several theories for why unemployment might be narrowly affecting the youngest workers. The most conventional explanation is that the labor market gradually weakened after the Federal Reserve raised interest rates. White-collar companies that expanded like crazy during the pandemic years have slowed down hiring, and this snapback has made it harder for many young people to grab that first rung on the career ladder.

But another explanation is too tantalizing to ignore: What if it’s AI? Tools like ChatGPT aren’t yet super-intelligent. But they are super nimble at reading, synthesizing, looking stuff up, and producing reports—precisely the sort of things that twentysomethings do right out of school. As I wrote:

Reenvisioning the college major

Assuming the requirement for students to complete a major in order to earn a degree, colleges can also allow students to bundle smaller modules – such as variable-credit minors, certificates or course sequences – into a customizable, modular major.

This lets students, guided by advisers, assemble a degree that fits their interests and goals while drawing from multiple disciplines. A few project-based courses can tie everything together and provide context.

Such a model wouldn’t undermine existing majors where demand is strong. For others, where demand for the major is declining, a flexible structure would strengthen enrollment, preserve faculty expertise rather than eliminate it, attract a growing number of nontraditional students who bring to campus previously earned credentials, and address the financial bottom line by rightsizing curriculum in alignment with student demand.

Uber: The Evolution of Uber’s Search Platform

Shopify: Introducing Roast – Structured AI Workflows Made Easy

Sem Sinchenko: Why Apache Spark is often considered as slow

Meta: Collective Wisdom of Models: Advanced Feature Importance Techniques at Meta

We all love our little language wars: the vitriol, the expletives, the dank memes that accompany them.

You may recall, as your friendly neighborhood Anonymous Rust Dev, that I’ve been magnanimous and kind to the Python language. This is because I’m above it all – I can find the beauty in any language if I look hard enough, and I’m willing to give credit where it’s due.

But even I have my breaking point, which I’d like to explore today. How clunky does a language need to be to make itself my foe?

1) Avoid Being The Only Engineer On A Team Early On In Your Career

2) Hype Won’t Fix Your Business Problems

3) Solving Business Problems Is Hard

4) Don’t Skip The Fundamentals

5) Don’t Be Above Putting in the Work

6) Ask Why – But Don’t Always Expect A Great Answer

7) Learn to Communicate with Non-Data People

What’s Next? Predictions for the Future of Search and SEO

• Visibility Shifts from Keywords to Topical Coverage
• Personalization Becomes Standard
• Paid Placements and Monetization in AI Mode
• The Agency Divide: Traditional vs. Modern Approaches
• New Analytics and Search Console Features
• Final Thought: Search Is Now a Conversation

As AI Mode evolves from an experiment to a default experience, search will feel less like a transaction and more like an ongoing conversation. Success will go to brands that can be present, relevant, and trusted at every step—no matter how the questions change.

To grow and thrive in the rapidly evolving AI landscape, organizations must strategically invest in their data engineering capabilities.

In today’s modern digital landscape, businesses are generating heavy data daily which can be processed, analyzed and interpreted for future scalability and growth. This is when AI-driven systems become integral across industries to help create real-time analytics, forecasting and initiating AI-driven automation. Beverly D’Souza, a Data Engineer at Patreon (previously worked at Meta) has played a key role in improving data workflows, processing data at pace and launching machine learning models. Having experience with ETL pipelines, cloud data systems, and AI analytics, she shared, “Building scalable AI-powered data pipelines comes with key challenges and to overcome these obstacles, organizations must implement distributed computing frameworks that can handle large-scale data processing efficiently. Incorporating AI-driven automation helps streamline data processing tasks, making the entire system faster and more efficient.”

The Efficacy of Standardized Interfaces in Reducing Manual Intervention, Improving System Resilience, and Enabling Advanced Governance in Enterprise Data Platforms.

Future Outlook
Agentic AI, MCP, and A2A are poised to reshape enterprise data platforms in the next 3–5 years. We will likely see middleware for agents become mainstream: platform vendors already integrate these protocols. Systems will evolve from static ETL pipelines into adaptive, self-optimizing networks. For instance, future data lakes might auto-tune storage tiers based on usage patterns discovered by agents, or data warehouses could self-partition hot tables. As Microsoft’s announcements highlight, AI is moving toward an “active digital workforce” [33]: think LLMs that don’t just suggest queries, but execute workflows end-to-end. With A2A, agents from different vendors and clouds will interoperate, breaking current silos. Enterprises will embed AI into governance: agentic systems continuously audit for compliance.

We may also see advances in model capabilities driving agentic efficiency — e.g., hybrid systems where a symbolic planner guides LLMs, or LLMs with built-in code execution (like Azure’s CUA) making some MCP calls redundant. Standards (MCP, A2A) will likely expand; Google’s A2A is already collaboration with 50+ partners [40], promising broader interoperability. In short, the data platform of the future could sense, reason, and act: as data patterns shift, agents reconfigure pipelines; when costs spike, agents throttle resources; when new regulations arrive, agents update data handling policies. This vision of an adaptive, self-driving data platform is on the horizon thanks to agentic AI and these new protocols.