The Establishment Gets Disrupted — From Within
Yann LeCun built some of the foundations that modern deep learning runs on. For years he ran AI research at Meta, watching the field he helped create race toward systems he believes are fundamentally flawed. Late last year he left. Today, his one-month-old startup, Advanced Machine Intelligence Labs (AMI Labs), announced it has raised over $1 billion in seed funding at a $3.5 billion valuation.
Twelve employees. One month old. Three and a half billion dollars.
The investors include Jeff Bezos, Mark Cuban, and a collection of VC firms betting that LeCun's alternative approach — moving away from the transformer-based paradigm that powers ChatGPT, Gemini, and Claude — could be the path forward that the rest of the industry is missing. LeCun has been vocal for years that current large language models are hitting a ceiling: impressive at language, but fundamentally limited as a path to general intelligence.
Whether he's right is an open question. But the fact that investors are willing to write billion-dollar checks to find out says a lot about where we are in 2026.
GPT-5.4: A Million Tokens and a Mission Statement
Meanwhile, the establishment isn't standing still.
On March 5th, OpenAI released GPT-5.4, which it's billing as its "most capable and efficient frontier model for professional work." The headline stat: a 1-million-token context window in the API — roughly 50–100× longer than previous generations. That's not just a benchmark number. It means GPT-5.4 can ingest an entire codebase, a year of company documents, or a full legal case file in a single prompt.
Pair that with mid-response planning (the model can now think through steps while generating output), and you have something that starts to look genuinely useful for complex, long-running enterprise workflows — not just quick Q&A.
OpenAI also launched ChatGPT for Excel on the same day, embedding GPT-5.4 directly into spreadsheets. The vision: describe what you want in natural language, and the model builds the formulas, models, and visualisations. For analysts who spend their days in Excel, this is either very exciting or mildly terrifying, depending on your job security.
Google Quietly Ships Gemini 3.1 Flash-Lite
While GPT-5.4 grabbed headlines, Google rolled out Gemini 3.1 Flash-Lite — a faster, cheaper variant of its Gemini 3.1 model family. No splashy press conference, just a model that costs less to run and returns answers quicker.
This is Google's play at the other end of the spectrum: not raw capability, but economics and speed. For developers building real products that need to handle millions of requests cheaply, Flash-Lite is worth paying attention to. Google also shipped a wave of Gemini updates across Workspace — Docs, Sheets, Slides and Drive all got deeper AI integration this week.
The theme is clear: Google is no longer just doing AI demos. It's embedding Gemini into every surface millions of people use every day.
NVIDIA's Hardware and Open-Source Moves
All of this AI capability needs hardware to run on. NVIDIA is making sure it controls that layer too.
Two moves this week: first, an announcement of the Rubin supercomputer platform — the next-generation hardware stack designed to power the kind of massive AI workloads that come next. Second, and arguably more interesting for developers: NemoClaw, an open-source AI agent platform aimed squarely at enterprise adoption.
Open-sourcing agent infrastructure is a smart play. It lets NVIDIA become the default plumbing for enterprise AI without needing to win every model race. Get the tooling layer locked in, and the GPU sales follow.
NVIDIA's GTC 2026 conference kicks off next week (March 16–20), where more announcements are expected — particularly around healthcare AI and physical AI systems.
Robots Are Getting Real
On the physical AI front: Figure AI's Helix 02 humanoid robot — backed by NVIDIA — had a notable public demonstration this week, showcasing autonomous capabilities that drew both admiration and the predictable Elon Musk commentary.
Humanoid robotics remains one of the longer bets in tech, but the cadence of credible demos is accelerating. When NVIDIA starts backing specific robot companies, it's worth tracking which ones.
The Macro Picture: $2.52 Trillion and a Wave of Structural Layoffs
Zoom out and the numbers are staggering. Gartner projects worldwide AI spending will hit $2.52 trillion in 2026. That's not just GPU purchases and API calls — it includes the consulting, the tooling, the implementation, and the people trying to make all of it actually work in businesses.
But this week also brought a less comfortable story: a notable wave of layoffs across the tech industry — and unlike previous rounds, these aren't trimming bloat from the pandemic hiring surge. These are structural layoffs, roles that AI is beginning to automate at a meaningful scale. Senior enough positions to matter. Early enough in 2026 to be a signal.
The counter-argument is always job creation — and historically, technology does create more roles than it displaces. But the pace and specificity of current AI capabilities is different from previous waves of automation. It's worth watching whether the job creation keeps up.
What This Week Tells Us
A few things are becoming clear in early 2026:
The model race is fragmenting. There's no longer a single "best model." GPT-5.4, Gemini 3.1, Claude's latest, Qwen 3.5 from Alibaba — different models are winning on different tasks, at different price points. For builders, this is good. For anyone trying to make a simple "which AI should I use?" recommendation, it's increasingly complicated.
The hardware layer is converging on NVIDIA. LeCun's billion-dollar bet is partly a hardware bet too — different architectures will require different silicon. But for now, NVIDIA owns the stack.
The money is still flowing, despite bubble warnings. AMI Labs at $3.5B with twelve employees is either a sign of irrational exuberance or a sign that investors understand that the window to back foundational AI research is closing fast. Probably some of both.
The impact on work is arriving, not theoretical. The layoffs this week aren't scare stories — they're data points. The question isn't whether AI will change professional work. It's how fast, and whether the new roles emerge quickly enough to absorb the disruption.
One more week in the fastest-moving field in the history of technology. See you on the other side of the next one.
Sources: New York Times, devFlokers AI Digest, OpenTools.ai, Reuters, Google Workspace Blog, NVIDIA Blog
Top comments (0)