Computerspeak from Alexandru Voica

A weekly newsletter from @alexvoica focused on AI and emerging technologies.

We share the most important stories in AI daily and weekly recaps 👇

[Alexandru Voica 💀 @alexvoica]

https://t.co/VeWrrSIKsc

[Alexandru Voica 💀 @alexvoica]

https://t.co/ADPraweze1

[Alexandru Voica 💀 @alexvoica]

The deeper issue is that compute was never a reliable proxy for risk in the first place. It measures how much energy went into training a model, not what the model can actually do or what harms it might cause. A serious classification framework should look at capabilities, deployment context, and real-world impact.

Many of models won't pose anything close to systemic risk (e.g. https://t.co/9lPzZPELjj), and some of the most capable and potentially risky systems (especially from China) may be trained more efficiently through distillation and fall below the line entirely.

The threshold, in other words, is becoming both overinclusive and underinclusive at the same time.

It's time to move past compute thresholds altogether and toward evaluation methods that reflect what we in Europe actually should care about: what these models can do, how they're used, and where the genuine risks lie.

That shift would produce better regulation, more proportionate obligations, and a framework that doesn't need emergency patching every time the industry takes another step forward.

Unfortunately, many in the AI Office are too set in their ways (or too influenced by doomers) so they'll never make this change.

[Alexandru Voica 💀 @alexvoica]

In a call with reporters ahead of the release of GPT-5, @sama described "a model that continuously learns as it's deployed from the new things it finds" something that to him "feels like AGI."

In this week's Computerspeak, I attempt a more detailed spec for what AGI (or superintelligence) could mean:

🧠 Performs at or above the equivalent of a model with 100 trillion parameters
🔌 Requires no more than 1 kilowatt of power to operate
🧑‍💻 Is a world model that's more than just a pretty video generator

https://t.co/7j1SxV8c1v

[Alexandru Voica 💀 @alexvoica]

We have reached AGI.

[Alexandru Voica 💀 @alexvoica]

While I’m pretty sure that the Venn diagram of things that @_KarenHao and executives from the AI industry agree on is generally small, this week we saw it getting a tad wider in real time.

In her new book Empire of AI, she explains that AI boomers and doomers are two sides of the same coin because both groups fervently believe in two ideas: AGI is inevitable and it is coming soon. The only difference is that the former imagines a utopia while the latter fears a full-on annihilation.

Nvidia CEO Jensen Huang poured some much needed cold water on the boomers this week, arguing that they are fostering fear to advance their own agenda.

My thoughts on this public spat in this week's newsletter 👇

https://t.co/udWJ9Fh5mv

[Alexandru Voica 💀 @alexvoica]

I spoke to @SPGMarketIntel about Europe's ambitious plan to invest €200 billion in AI gigafactories:

https://t.co/QOYX18DagH

[Alexandru Voica 💀 @alexvoica]

World models are essential testing grounds for AI agents.

Researchers from @MBZUAI are working on PAN, a world model that can simulate infinitely diverse realities, from simple physical interactions to complex multi-agent systems: https://t.co/3arkuyiuIF

[Alexandru Voica 💀 @alexvoica]

EVERYTHING IS COMPUTER! #GoogleIO

[Alexandru Voica 💀 @alexvoica]

After attending a discussion about AI and copyright at the Palace of Westminster on Tuesday, here are two proposals that could help the UK to develop a progressive, growth-focused, and standards-based approach to responsible AI:

https://t.co/idQ4MFYmoA

[Alexandru Voica 💀 @alexvoica]

A new study from @pewresearch reveals a significant divergence between experts and the general public regarding the benefits of AI.

This trust gap is completely understandable, if we look at what happened during the previous wave of automation:

https://t.co/5L4shSAmw1

[Alexandru Voica 💀 @alexvoica]

In this week's @computerspeak_ newsletter, I explain some of the reasons why startups have historically found it hard to rely on EU supercomputers for AI development:

https://t.co/7dWrpwb9Mx

[Alexandru Voica 💀 @alexvoica]

Today's AI ecosystem has been built on top of a compute and data infrastructure created during the golden age of the technology industry.

Here's how the coming trade war between the United States and China will affect that:

https://t.co/xq6CZhcIxe

[Alexandru Voica 💀 @alexvoica]

A lot has been written about the Make It Fair campaign launched by representatives of the UK creative industry earlier this week.

I look at the solutions proposed by @jujulemons, @imogenheap, and @johnthornhillft:

https://t.co/tld2S2gCev

[Alexandru Voica 💀 @alexvoica]

Elon Musk recently said that superintelligence will be among the top 10 most important milestones in the history of life on Earth.

I explain the consequences of his statement in this week's @computerspeak_ newsletter:

https://t.co/9TjqA6ECy5

[Computerspeak from Alexandru Voica @computerspeak_]

@alexvoica 👋

[Alexandru Voica 💀 @alexvoica]

For more than two years, we’ve had to endure the droning of a vocal group who have hijacked the public discourse around generative AI.

Thankfully, at the #AIActionSummit in Paris, we've finally said adieu et à bientôt to most of them:

https://t.co/VsdE3QoV5V

[Alexandru Voica 💀 @alexvoica]

In this week's @computerspeak_ newsletter, I look at four developments from the world of reasoning models, and how they challenge the established narrative around AI agents:

1️⃣ @GitHub’s CEO @ashtom announcement of "agent mode," a new Copilot feature that elevates AI pair programming to peer programming.

2⃣ @fb_engineering's new ACH tool for enhanced software testing (which builds on years of great research from @Mark_Harman and his team in search-based software engineering)

3⃣ @huggingface's paper asking for AI companies to reconsider the development of fully autonomous AI agents, presented by @mmitchell_ai at #IASEAI25

4⃣ An excellent article from @joshgans reflecting on the process of authoring a published paper with the help of OpenAI's o1 model - and what happens if research becomes "cheaper than search."

https://t.co/4a6ZeNegUa

[Alexandru Voica 💀 @alexvoica]

When @deepseek_ai released R1, a well-performing open weights reasoning model, the response from many in the American media and industry was to frame it as the latest salvo in the US-China "AI war."

For @computerspeak_, my AI avatar explains why that's unhelpful 👇

[Alexandru Voica 💀 @alexvoica]

This was a productive week for people keen to share cringe geopolitical takes on X, many of whom positioned @deepseek_ai's new R1 model as a "Sputnik moment" in a zero-sum war between the United States and China:

https://t.co/GJEknInwQQ