Lightnews — Scholar-powered news

Peter Bloem @pbloem.sigmoid.social.ap.brid.gy · 15d

me: Maybe I should consider the environmental impact of my coffee consumption.

tech CEOs: New strategy! Time to landfill a quarter million perfectly viable computers we already built.

https://bitbang.social/@NanoRaptor/115258704574406678

Nanoraptor (@[email protected])

Attached: 1 image The NeXTlevel was NeXT Software's first handheld device, eagerly awaited but canned just two weeks from release when Apple purchased NeXT and chose not to pursue gaming. All 250,000 release devices were pulled from channels and unceremoniously rolled into landfill out northwest of Dubbo.

bitbang.social

Reposted by Peter Bloem

Misha @mishavelthuis.social.edu.nl.ap.brid.gy · 16d

Triple rainbow at #sciencepark #amsterdam.

Campus management and the universe joining forces in the fight against hatred and bigotry.

#universityofamsterdam #uva

Picture showing the bridge to Science Park (University of Amsterdam), which is painted in rainbow colors. In the background you can see a real rainbow (which like all rainbows, is a double rainbow), which put together creates a triple rainbow.

1

Peter Bloem @pbloem.sigmoid.social.ap.brid.gy · 19d

Holy hell, GPT-5 is snappy today. Did they switch to a diffusion model?

Peter Bloem @pbloem.sigmoid.social.ap.brid.gy · 19d

@djoerd That's the other explanation. It's certainly true for things like stock prediction.

It's probably an element of both. Good forecasting is probably more about recognizing when you don't have an edge and then defaulting to a simple baseline.

Still, the fact that certain humans are […]

Original post on sigmoid.social

sigmoid.social

1

Peter Bloem @pbloem.sigmoid.social.ap.brid.gy · 19d

I wonder how close this comes to the "committee prompting" I made up for this:

https://profane-tmesis.info/chain-of-thought

Chain of thought

A: Our starting assumption is that the message is toxic. What do you think? B: Context is key. The message reads like a joke. It may be satirical. C: We have limited access to context. However, the fact that it is a joke does not preclude the possibility of it being toxic.

profane-tmesis.info

Peter Bloem @pbloem.sigmoid.social.ap.brid.gy · 19d

It seems like this is mostly a hodge-podge of different API calls and various prompts.

Still, gives the lie to the theory that it's all just "regurgitating training data", "unable to come up with novelty", etc […]

Original post on sigmoid.social

sigmoid.social

2

Reposted by Peter Bloem

George Monbiot @georgemonbiot.bsky.social · 19d

Power is the rock on which truth founders. The mainstream media, with a few exceptions, is a single-issue lobby group, whose purpose is to assert the rights of capital.
Here's what I've learnt about the industry in 40 years in journalism.
www.theguardian.com/commentisfre...

I have now been a journalist for 40 years. The forces ranged against my profession have never been so powerful | George Monbiot

Once, I believed that humanity’s problem was an information deficit. Now, I know you can’t speak truth to power if power controls your words, says Guardian columnist George Monbiot

www.theguardian.com

67 800 1.8K

Peter Bloem @pbloem.sigmoid.social.ap.brid.gy · 19d

It's like taking a cruel and sociopathic kid and mocking him for him weak muscles. You may not like it when he comes back from the gym.

Peter Bloem @pbloem.sigmoid.social.ap.brid.gy · 20d

On a higher level, this shows again that it's a real mistake to lean to heavily on supposed AI inabilities in your AI criticism.

It's true that we can't do it yet, but it's not true that it's impossible. Doing this accurately is very much within reach. We should be worrying about whether it […]

Original post on sigmoid.social

sigmoid.social

1

Peter Bloem @pbloem.sigmoid.social.ap.brid.gy · 20d

One reason this hasn't been pursued more enthusiastically (I'm guessing) is cost. This kind of feedback loop requires long contexts and many image generation runs, which takes too much time and compute.

Another is probably that really accurate image generation has a nasty flipside in generating […]

Original post on sigmoid.social

sigmoid.social

1

Peter Bloem @pbloem.sigmoid.social.ap.brid.gy · 20d

It's also one of those things that is almost on the way out.

You can easily create a feedback loop between the generator and the LLM (I think Gemini is experimenting with this)? Eventually, this should lead to a coherent image: the chatbot has the knowledge and can pretty consistently apply it […]

Original post on sigmoid.social

sigmoid.social

1

Peter Bloem @pbloem.sigmoid.social.ap.brid.gy · 20d

Chatbots like ChatGPT (for now) a pretty fundamentallty different technology than the sub-module they call on to generate an image. It works differently, it's trained differently and it's subject to very different levels of guardrailing and fintuning.

To be fair, this is mostly on OpenAI. If […]

Original post on sigmoid.social

sigmoid.social

1

Peter Bloem @pbloem.sigmoid.social.ap.brid.gy · 20d

A thread by fediverse favorites the Vagina museum on AI-generated inaccuracies.

They're mostly right to highlight this, but it shows an issue in the way people reason about AI's reliability: they look at the dream-logic nonsense that comes out of image generators and generalize it to the rest […]

Original post on sigmoid.social

sigmoid.social

1

Peter Bloem @pbloem.sigmoid.social.ap.brid.gy · 22d

@leonoverweel That's a fair point. But I was mostly thinking of the general population of people who don't (or rarely) directly use modern AI.

They will see an increase of slop and not much else. The benefits of RAG and reasoning are on the horizon for them, but not directly noticeable yet […]

Original post on sigmoid.social

sigmoid.social

Peter Bloem @pbloem.sigmoid.social.ap.brid.gy · 22d

@anandphilipc I think this is why all the AI companies are investing in their own search engines and in licensing agreements.

Maybe the open aource alternative is to do RAG on scihub. Perhaps we also need to push decentralized web search more (it exists but it isn't very mature yet).

1

Peter Bloem @pbloem.sigmoid.social.ap.brid.gy · 22d

The problem is that this only works if RAG can source high-quality, human-made sources of information. At the moment, GPT-5 at least seems well equipped to ignore the slop, but that there still needs to _be_ something other than slop.

We still need a Wikipedia that is entirely manually built […]

Original post on sigmoid.social

sigmoid.social

1

Peter Bloem @pbloem.sigmoid.social.ap.brid.gy · 22d

This feels counter-intuitive when most people currently experience AI as a wave of automatically generated bullshit, but if you look at the latest generation of models, with long reasoning and good RAG, they ave a surprisingly high hit rate for accurately doing and summarizing literature […]

Original post on sigmoid.social

sigmoid.social

2

Peter Bloem @pbloem.sigmoid.social.ap.brid.gy · 22d

I know it's hopelessly naive of me (and it's going to go over like a lead balloon on here), but I often think that AI can kind of reset the scales on Brandolini's law.

You see some of this dynamic with Grok already, where despite Musk's best efforts, the poor bot just keeps outputting consensus […]

Original post on sigmoid.social

sigmoid.social

1 1

Reposted by Peter Bloem

-Boulet- @bouletcorp2.mastodon.social.ap.brid.gy · 23d

Rogatons Rediffusions - 22/01/2024 - "Gratouilles"
Source: https://www.bouletcorp.com/rogatons/2024/01/22

LES LIVRES: https://www.exemplaire-editions.fr/exemplaire/libraire/boutique

Gratouilles
Visite chez ma sœur
Boulet est sur son téléphone, assis dans un canapé, l'air blasé.
Un petit "froosh froosh" se fait entendre.
Boulet tend la main pour caresser un chat distraitement. Il lui dit "Ouais ouais, bon pépère."
Plus tard, Boulet discute avec sa sœur à table, en buvant un café. Un froosh-froosh se fait entendre. Il tend la main pour caresser le chat en faisant "Oui oui pépère."
Sœur: "Tu vois ? Il est gentil."

La nuit...
Boulet est dans un canapé-lit et dort. Un froosh-froosh se fait entendre.
Boulet tend la main dans un demi-sommeil pour caresser le chat en marmonnant "mgnff'm pépère."

Gare de Nancy. Boulet est sur son téléphone, son sac de voyage à côté de lui. Un froosh-froosh se fait entendre. Boulet tend distraitement la main en disant "ouais ouais bon pépère..." et touche la tête d'un pigeon.

Boulet sursaute quand il est recouvert par un nuage de pigeons en panique, sous l'œil surpris d'une vieille dame.
La vieille dame: "Il faut pas toucher aux pigeons."
Boulet (débraillé et encore sous le choc): "Je sais."
La vieille dame: "Ils aiment pas qu'on les touche, les pigeons."
Boulet: "Je. Sais."

11 3

Reposted by Peter Bloem

Shine McShine 🐌 @shine-mcshine.neopaquita.es.ap.brid.gy · Sep 3

cuando me preguntan que por qué no me voy a bluesky

Meme de dos paneles de la película Rogue One. En el panel superior el Director Krennic, rodeado de soldados de asalto imperiales, pregunta "Mastodon? Really? Man of your talents?" y en el panel inferior Galen Erso contesta "It's a peaceful life"

1 110 32

Peter Bloem @pbloem.sigmoid.social.ap.brid.gy · 28d

Can't believe they found an even bigger asshat than Musk to be the new richest person.

2

Reposted by Peter Bloem

Bruce Lawson ✅ ♫ ♿ ✌️♂️✊ @brucelawson.vivaldi.net.ap.brid.gy · Sep 9

Apparently, it's bad luck to mention a "MacBook" in a software company. You have to call it "the Scottish laptop".

14 120 93

Peter Bloem @pbloem.sigmoid.social.ap.brid.gy · Sep 9

I think the Comma approach is the way forward for (more) ethical AI, but this is a very useful model to have in the mean time. The instruction tuning in particular is sorely missed in Comma (although they're working on it). And the larger model size puts it a little closer to being actually useful.

Peter Bloem @pbloem.sigmoid.social.ap.brid.gy · Sep 9

New open model: Apertus

https://www.swiss-ai.org/apertus

Tech report here:

https://github.com/swiss-ai/apertus-tech-report/blob/main/Apertus_Tech_Report.pdf

They release 8B and 170B models, pre-trained and instruction tuned. Open weight and open data.

Data ethics are somewhere between OLMo […]

Original post on sigmoid.social

sigmoid.social

1

Reposted by Peter Bloem

Emelia 👸🏻 @thisismissem.hachyderm.io.ap.brid.gy · Sep 6

Recently there has been a lot of discourse about ActivityPub and AT Protocol which has been quite dividing and heated.

Yesterday at the Social Web CG meeting (the group that maintains the ActivityPub and related specifications), I proposed releasing a statement that counters the narrative that […]

Original post on hachyderm.io

hachyderm.io

7 78 13