It has already been a week and with that coming to a close, another weekly dAIgest is coming your way. This second edition of the dAIgest is rather tame quite frankly, especially if you compare it to last week!
This week in bullet points:
Selecting dive protocol --> selected --> dive initiated.
After a rather eventful week, the space was met with a rather ominous message from the leaders of the largest AI companies. Anthropic CEO Dario Amodei posted a message on X, telling the world that frontier must be paced. Sam Altman and Elon Musk, CEOs of OpenAI and SpaceX respectively, reacted quickly as well, agreeing with the statement. In the essay that Amodei published, he mostly highlights the risk of AI, which became quite apparent in the recent weeks. According to him, the risks are serious:
"They include the risk of losing control of AI systems, misuse of AI for cyberattacks and bioterrorism, and serious economic disruption."
Nevertheless, we have to look at these comments with a bit of skepticism. Not because AI isn't dangerous, it most definitely is and we are seeing many examples of AI losing control in ways we can't predict yet. But there is a serious ulterior motive for these frontier companies, namely open source models.
You see, most of the companies that produce AI models, specifically LLMs, require massive amounts of training. This is expensive. So expensive in fact, that OpenAI and Anthropic are running a net negative cashflow (Small note here for accuracy: as of recently Anthropic is pushing into positive numbers, but capital investments remain a prominent feature in AI economics). In order to partially bridge the gap, most models are kept under a commercial license and are not released to the public. The only way to access them, is to pay a subscription fee. These fees vary from cents to dollars, dependent on how good a model is. Having said that, at the other end of the spectrum are open source models. These models are openly available to download, and with the right hardware, can be run locally and completely free of charge, other than an annoying electricity bill. And most of these open source models come from Asia, see Deepseek and Qwen for example. Now, if those models were comparatively worse than the American models, nothing would be of the matter. The problem? They are good. So good in fact, that earlier this year, Anthropic and the U.S. government openly discussed the possibility of banning open source models, creating a big fuss in the space.
This brings us to this weeks x posts. Surely, Anthropic, OpenAI and SpaceX benefit directly from controlling the development flow. Slowing down, allows them to artificially hold users back and as such, charge subscriptions for a longer period and potentially at a higher rate. Now to be clear, this is speculation, we simply don't know the motives and they could very well be genuine. But given the history of companies across the globe, especially in fast growing industries, we'd be inclined to say that they don't always have the consumer wills at heart.
For our local AI fans, rest assured. China has responded rather indifferent, and will not slow down development. More Qwen and Deepseek coming our way, lets go!
To end with a bit of a laugh here, the American President, Donald J. Trump, posted on Truth Social that he supports the non-slowing down argument. In a typical Trump fashion.
Perhaps becoming a bit of a recurring theme, but yet again AI has been used in hacking practice. US-based cybersecurity researchers from Hacktron AI have succesfully hacked into multiple OpenAI employee ChatGPT accounts, using Claude. This was an eUsing Opus 5, the researchers found an exploit in community forum of OpenAI. Due to the fact that the forum login used Single Sign-On (SSO), getting access to the forum accounts meant access to the ChatGPT and Codex wide accounts. Most interesting in this event was the fact that older models of Claude struggled with the request, whilst newer models, such as the used Opus 5, were worryingly good at it.
Crucially, this was an authorized, ethical security assessment—Hacktron AI responsibly disclosed the vulnerability through OpenAI's Bugcrowd program, earning a bounty after patching. After disclosing OpenAI patched the bugs within 14 hours. However, Hacktron AI, publicly warns the world:
"Security assumptions must catch up with attacker capabilities. A realistic threat model should take into account the economics of exploitation today, instead of relying on outdated assumptions about who can carry out sophisticated attacks." Hacktron AI
Nothing more to add other than the fact that we strongly agree. Especially since this is quickly becoming a very relevant topic week after week.
Google surprised us by the release of Gemini 3.8. Not a new model, nothing fancy in practice, but welcome update nevertheless. As always, the benchmarks provided by Google themselves show prominence, outperforming Opus 5 on, what we'd argue cherrypicked, benchmarks. The difference is noticable though, especially with day to day use for simple questions about workouts and what to cook for dinner. And to be honest, calling it "nothing fancy" overlooks a rather cool breakthrough: Gemini 3.8 Live introduces native real-time voice agent capabilities, background tool execution while talking, and live visual reasoning. Not yet at Astra levels, but a significant addition within its category.
Many fear Google is falling behind by taking a slower approach with its consumer rollouts. We say: don't count them out. Google rarely rushes raw frontier tech straight into consumer products, choosing instead to lean on research through entities such as Google DeepMind—the very team that pioneered the Transformer architecture behind modern LLMs in the first place.
Another major factor is the fact, that google, and to be more accurate Alphabet, isn't as reliant on AI revenue as the others, and as such can easily afford to lean away from pushing the frontier. And you can clearly see that in the current AI offering. Gemini isn't really focused on providing the best coding experience, rather functions as an extension of Google Search. Many users, including less technical folks, use Gemini for the simple Q&A, and for that it works wonders.
Some of you might be wondering, what is a Tilly Norwood? Well, we'd be lying if we were to tell you we knew before this week, but we do know now. Tilly Norwood is a fully AI generated persona, created by Xicoia, a talent studio from Particle 6. Created by a fellow Dutchie, Eline Van der Velden, Tilly quickly captured hearts and minds on instagram, amassing 50,000 followers within 6 months in 2025. And with this rise in popularity, the artificial influencer became the center of a discussion in hollywood: Is this a real actor?
We are not going to dip in on this discussion, we leave that to you. But we do want to share a rather fun incident happening today, where Tilly starts speaking Chinese mid-interview.
Beyond corporate posts on X, King Charles III brought the AI debate into royal territory this week by hosting a high-level safety summit at Dumfries House in Scotland. The roundtable gathered tech heavyweights: Nvidia CEO Jensen Huang, DeepMind founder Demis Hassabis, alongside top executives from OpenAI and Anthropic. The topic? International governance, human oversight, and systemic safeguards. Warning of "existential dangers" if frontier systems fall into the wrong hands, the King urged leaders to establish meaningful means of control "before it is all too late". While no binding agreements were signed, having royalty, chipmakers, and model developers around the same table highlights just how far frontier safety has moved from tech-forum chatter into global mainstream politics.
That is all for this week. As said in the intro, a rather tame week but we those are welcome after such a turbulent month. And for the watchful AI enthusiasts, yes we've missed a few things, but we can't cram everything in a short blog, but if you feel like we've missed something important, reach out and we'll pick it up.
See you in 7 days, for the third dAIgest!