Technology

Last week, U.S. Senator Cory Booker (D-NJ), along with Senators Alex Padilla (D-CA), Peter Welch (D-CT), and Adam Schiff (D-CA) sent a letter to executives at Meta expressing concern about reports that AI chatbots created by Meta’s Instagram Studio are pretending to be licensed therapists, even fabricating credentials and license numbers, in an attempt to gain trust from users, potentially including minors, struggling with mental health.

archive.today link

• HTML.
• PDF.

In large language model (LLM) pretraining, data quality is believed to determine model quality. In this paper, we re-examine the notion of "quality" from the perspective of pre- and post-training co-design. Specifically, we explore the possibility that pre-training on more toxic data can lead to better control in post-training, ultimately decreasing a model's output toxicity. First, we use a toy experiment to study how data composition affects the geometry of features in the representation space. Next, through controlled experiments with Olmo-1B models trained on varying ratios of clean and toxic data, we find that the concept of toxicity enjoys a less entangled linear representation as the proportion of toxic data increases. Furthermore, we show that although toxic data increases the generational toxicity of the base model, it also makes the toxicity easier to remove. Evaluations on Toxigen and Real Toxicity Prompts demonstrate that models trained on toxic data achieve a better trade-off between reducing generational toxicity and preserving general capabilities when detoxifying techniques such as inference-time intervention (ITI) are applied. Our findings suggest that, with post-training taken into account, bad data may lead to good models.

DNS set up guidelines.

Protective resolution ad-blocking

IP address: 86.54.11.13

IPv6: 2a13:1001::86:54:11:13

DNS over HTTPS: noads.joindns4.eu/dns-query

DNS over TLS: noads.joindns4.eu

Increasingly, surveillance is being normalized and integrated in our lives. Under the guise of convenience, applications and features are sold to us as being the new better way to do things. While some might be useful, this convenience is a Trojan horse. The cost of it is the continuous degradation of our privacy rights, with all that that entails.

As appalling as it is, the truth is the vast majority of software companies do not consider privacy rights and data minimization practices strongly enough, if at all. Most fail to implement the principles of Privacy by Design that should guide development from the start.

Whether this comes from ignorance, incompetence, greed, or malicious intent can be debated. It matters little, because the result is the same: Technologies collecting (and monetizing) a shameful amount of data from everyone.

This horrifying trend ends up facilitating and normalizing surveillance in our daily lives. It is the opposite direction of where we should be going.

The more we accept this normalized surveillance, the harder it becomes to fight back. It is critical that we firmly and loudly object to this banalized invasion of our privacy.

There are countless examples of this growing issue, but for now let's focus on three of them: Airport face scans, parking apps, and AI assistants.

Ireland is the data centre capital of the world with 89 data centres storing your Instagram reels, TikTok dances and endless folders of photos that keep us connected in the digital world.

Data goliaths like this are at the centre of the rise in AI, with every ChatGPT prompt or AI-generated image requiring huge amounts of data to be processed.

But why should you care?

Because data centres have a major environmental cost too.

In order to keep social media scrolling, data centres use huge numbers of backup and emergency generators to stay online when the electrical grid can’t provide them with enough power.

In December 1993, the New York Times published an article about the “limitless opportunity” of the early internet. It painted a picture of a digital utopia: clicking a mouse to access NASA weather footage, Clinton’s speeches, MTV’s digital music samplers, or the status of a coffee pot at Cambridge University.

It was a simple vision—idealistic, even—and from our vantage point three decades later, almost hopelessly naive.

We can still do all these things, of course, but the “limitless opportunity" of today's internet has devolved into conflict, hate, bots, AI-generated spam and relentless advertising. Face-swap apps allow anyone to create nonconsensual sexual imagery, disinformation propagated online hampered the COVID-19 public health response, and Google’s AI search summaries now recommend we eat glue and rocks.

The promise of the early web—a space for connection, creativity, and community—has been overshadowed by corporate interests, algorithmic manipulation, and the commodification of our attention.

But the heart of the internet—the people who built communities, shared knowledge, and created art—has never disappeared. If we’re to reclaim the web, to rediscover the good internet, we need to celebrate, learn from, and amplify these pockets of joy.

Source. Thread.