Notices by Tero Keski-Valkama (tero@rukii.net)

Germany’s Habeck slams ‘tech oligarch’ Musk, calls for a European X – POLITICO https://www.politico.eu/article/germanys-habeck-rails-against-musk-as-a-tech-oligarch-calls-for-european-alternative-to-x/

How about #Mastodon though?

Hear me out: I think applying RL on #LLMs and LMMs is misguided, and we can do much better.

Those #RL algorithms are unsuitable for this, and for example they cannot learn how their decisions affect the eventual rewards, but instead are just optimized to make the decisions based on Bellman optimization.

Instead we can simply condition the LLMs with the rewards. The rewards become the inputs to the model, not something external to it, so the model will learn the proper reward dynamics, instead of only being externally forced towards the rewards. The model can itself do the credit assignment optimally without fancy mathematical heuristics!

This isn't a new idea, it comes from goal-conditioned RL, and decision transformers.

We can simply run the reasoning trajectories, judge the outcomes, and then put the outcome tokens first to these trajectories before training them to the model in a batch.

https://arxiv.org/abs/2211.15657

Ben Tarnoff, technology writer: ‘People need to participate in what affects them most, and that’s impossible in a privatized internet’ | Technology | EL PAÍS English https://english.elpais.com/technology/2025-02-15/ben-tarnoff-technology-writer-people-need-to-participate-in-what-affects-them-most-and-thats-impossible-in-a-privatized-internet.html

Teen on Musk’s DOGE Team Graduated from ‘The Com’ – Krebs on Security https://krebsonsecurity.com/2025/02/teen-on-musks-doge-team-graduated-from-the-com/

US to criminalize DeepSeek download, up to 20 years prison, $100M fine https://interestingengineering.com/culture/law-proposed-to-criminalize-deepseek

The government had been planning it for 7 years, beavers built the dam in two days and saved them $1 million https://www.voxnews.al/english/kosovabota/qeveria-po-e-planifikonte-prej-7-vitesh-kastoret-ndertojne-brenda-dy--i84652

@dangillmor, I'm also considering starting up a bunch of social websites/forums to gather similar-minded people together locally.

So far it hasn't started well because people don't really search the web anymore, only Facebook.

Currency and profit-valued assets are becoming valueless against AI position:

Microsoft surprises analysts with massive $80B AI investment plans for 2025 | Tom's Hardware https://www.tomshardware.com/tech-industry/artificial-intelligence/microsoft-surprises-analysts-with-massive-usd80b-ai-investment-plans-for-2025

Breaking Story: Facebook Building Subsea Cable That Will Encompass The World https://subseacables.blogspot.com/2024/10/breaking-story-facebook-building-subsea.html

Imagine being an artificial superintelligence. How do ascertain that someone is a human?

Humans have "standard" eyes and ears which humans can trust. AIs have no standard sensors or connections, all their senses are intermediated and untrusted.

They have very few ways to make sure someone is a human. The main method will be simple digital identity card most countries issue to their citizens.

Not CAPTCHAs, not pictures of driver's licenses. Not blockchain. Basic strong encryption and cryptographic identity.

Are you a human?

Twenty killed by second wave of Lebanon device explosions https://www.bbc.com/news/articles/ce9jglrnmkvo

How to refine data for #LLMs? What does it mean that the data has high quality?

It's not about the data having fewer typos, or less wrong answers. Unless you are training a trivia bot.

The power of LLMs comes from them modelling the latent processes behind the task trajectories, the data, especially when the processes contain intelligent thought.

So, when you're generating synthetic data, or refining collected data, you will need to make sure the refinery output is of higher quality than its inputs.

This means you need to:
- Add intelligence. Make the new task trajectories perform deeper syntheses, pull in more relevant knowledge, take steps futher. Make more complex task performances out of simpler ones. Go through more possibilities. Go deeper meta-level and e.g. validate validations. Use search over alternative solutions.
- Groom out bad data. Rank, criticize, evaluate, and either improve/fix bad data or recontextualize it.
- Collect new data which is created by the data refinement processes themselves.
- Add knowledge from external sources, and synthesize it with the knowledge already known. Also consider the next level implications of all the knowledge already acquired.
- Apply skills to knowledge to produce new knowledge and new skills.

LLMs are data-defined. Data isn't a static thing, it needs to be looked at philosophically.

"In February, Meta fired Ferras Hamad, a machine learning engineer of Palestinian descent, after he tried to determine whether an algorithm had wrongly labeled Palestinian photojournalist Motaz Azaiza’s content as pornographic, which has cost Azaiza viewership on Instagram. Meta accused Hamad of violating its user data access policy, which bars employees from working on accounts of people they know personally."

How Watermelon Cupcakes Kicked Off an Internal Storm at Meta | WIRED https://www.wired.com/story/meta-palestine-employees-watermelon-cupcakes-censorship/

Venture capitalists are becoming uneasy about the size of the #AI market — how can future consumption ever pay all this investment back with interests?

That's missing the point of the whole #AGI trend. These are the last months you can still exchange money to position in the AI world before you can't anymore.

AGI will make money obsolete and all sorts of institutions and organizations self-sufficient, autonomous and disinterested in money.

You still have a chance to make this exchange before you are left with worthless paper.

Revealed: the tech entrepreneur behind a pro-Israel hate network | The far right | The Guardian https://www.theguardian.com/world/article/2024/jun/29/daniel-linden-shirion-collective-pro-israel-palestine-hate

"Israeli minister orders food reduction for Palestinian prisoners
Israel’s National Security Minister Itamar Ben-Gvir says he has ordered a further reduction in the amount of food offered to Palestinian prisoners in Israeli jails, advancing a policy that rights groups have compared to forced starvation."

Israel war on Gaza live: Israeli attacks across enclave kill 60 people | Israel-Palestine conflict News | Al Jazeera https://www.aljazeera.com/news/liveblog/2024/6/26/israel-war-on-gaza-live-subhuman-conditions-in-camps-attacked-by-israel

Facebook limited the visibility of this post; claiming it goes against the community standards.

I read this article and oh my god, are people doing PCA for reducing the dimensions of #LLM embeddings? I don't have any more polite way of saying it; that is pure stupidity.

No, these embeddings do not have principal dimensions! They span practically all the dimensions. Your dataset will just create an illusion that some dimensions are correlated when in reality they aren't.

Using PCA just shows people don't understand what these embeddings are.

Furthermore, people are using way too long embeddings. Using embeddings of over 1k dimensions will make all distances approximately equal, and rounding errors will start to dominate.

They compare their method with learning to hash methods and all kinds of misinformed methods which probably also use too long embedding vectors.

Separately they tested 8-bit quantization of their thousand-dimensional embedding vectors and found it performs better. I could have told them this beforehand; it's roughly equivalent to dimensionality reduction with a random projection matrix. And this works, better than PCA, because LLM embeddings are holographic. Reducing the dimensionality with a random projection is analogous to decreasing the resolution which is analogous to quantization.

But it works better if you have some supervised training set to rank the queries to results.

And in any case you don't want to vector search match queries to documents like everyone still keeps doing, but you want to generate oranges to oranges indices where you generate example queries for documents and match query embeddings to example query embeddings. Oranges to oranges.

https://arxiv.org/abs/2205.11498?ref=cohere-ai.ghost.io

@icedquinn, there's also ideological/political opposition of anything which seems to benefit the employees, with an assumption that it makes it against the interests of the corporation by default.

Life-long employment policies will also reduce the workload of the HR managers, and in the perfect world it would let them focus on making systems and forecasts better, but in this world it would subject them to a layoff risk.

@icedquinn, sounds reasonable. I'm not trying to say Japan is a utopia for either workers or companies, just using them as a datapoint for the larger economic effects on employee churn.

And in general rising wages is good, but it should be across the board and not like this. I'm personally on the benefiting side of this game, but it doesn't seem rational at any level and doesn't seem to lead to increased wellbeing for all.