vagabund

I don't why I would want to stare at vibecoded UI and an ad to relax

Yeah, stopped reading after the first paragraph. It's really so disrespectful to your audience.

Really fun to fly around, find my old apartment building etc.

It would be neat if you could drag and click to select an area to inpaint. Let's see everyone's new Penn Station designs!

Would guess it'd have to be BYOK but it works pretty well:

https://i.imgur.com/EmbzThl.jpeg

Much better than trying to inpaint directly on Google Earth data

More info: https://www.corememory.com/p/exclusive-openai-and-sam-altman...

I'd push back on a couple things here.

The notion that Scale AI's data is of secondary value to Wang seems wrong: data-labeling in the era of agentic RL is more sophisticated than the pejorative view of outsourcing mechanical turk work at slave wages to third world workers, it's about expert demonstrations and work flows, the shape of which are highly useful for deducing the sorts of RL environments frontier labs are using for post-training. This is likely the primary motivator.

> LLMs are pretty easy to make, lots of people know how to do it — you learn how in any CS program worth a damn.

This also doesn't cohere with my understanding. There's only a few hundred people in the world that can train competitive models at scale, and the process is laden with all sorts of technical tricks and trade secrets. It's what made the deepseek reports and results so surprising. I don't think the toy neural network one gets assigned to create in an undergrad course is a helpful comparison.

Relatedly, the idea that progress in ML is largely stochastic and so horizontal orgs are the only sensible structure seems like a weird conclusion to draw from the record. Saying Schmidhuber is a one hit wonder, or "The LLM paper was written basically entirely by folks for whom "Attention is All You Need" is their singular claim to fame" neglects a long history of foundational contributions in the case of the former, and misses the prolific contributions of Shazeer in the latter. Alec Radford is another notable omission as a consistent superstar researcher. To the point about organizational structure, OpenAI famously made concentrated bets contra the decentralized experimentation of Google and kicked off this whole race. Deepmind is significantly more hierarchical than Brain was and from comments by Pichai, that seemed like part of the motivation for the merger.

The huggingface spaces link doesn't work, fyi.

Sounds awesome in the demo page though.

> I highly doubt you are getting novel, high quality data.

Why wouldn't you? Presumably the end user would try their use case on the existing model, and if it performs well, wouldn't bother with the expense of setting up an RL environment specific to their task.

If it doesn't perform well, they do bother, and they have all the incentive in the world to get the verifier right -- which is not an extraordinarily sophisticated task if you're only using rules-based outcome rewards (as R1 and R1-Zero do)

My understanding is they built a performant suite of simulation tools from the ground up, and then they expose those tools via API to an "agent" that can compose them to accomplish the user's ask. It's probably less general than the prompt interface implies, but still seems incredibly useful.

Searching Spotify for user created playlists is still probably your best bet. Youtube has some good results too.

Here are two that might fit what you're looking for:

'90s K-pop: https://open.spotify.com/playlist/6mnmq7HC68SVXcW710LsG0?si=...

'00s minimal techno: https://open.spotify.com/playlist/6mnmq7HC68SVXcW710LsG0?si=...

There are sites to convert from spotify to another service if you don't have it.

No, but it's also biased toward their commercial partners. From this page [0], detailing their recommendation process:

> How do commercial considerations impact recommendations?

> [...] In some cases, commercial considerations, such as the cost of content or whether we can monetize it, may influence our recommendations. For example, Discovery Mode gives artists and labels the opportunity to identify songs that are a priority for them, and our system will add that signal to the algorithms that determine the content of personalized listening sessions. When an artist or label turns on Discovery Mode for a song, Spotify charges a commission on streams of that song in areas of the platform where Discovery Mode is active.

So Spotify's incentivized to coerce listening behavior towards contemporary artists that vaguely match your tastes, so they can collect the commission. This explains why it's essentially impossible to keep the algorithm in a historical era or genre -- even if well defined, and seeded with a playlist full of songs that fit the definition. It also explains why the "shuffle" button now defaults to "smart shuffle" so they can insert "recommended" (read: commission-generating) songs into your playlist.

[0]: https://www.spotify.com/ca-en/safetyandprivacy/understanding...

It may just be my perception, but I seem to have noticed this steering becoming a lot more heavy handed on Spotify.

If I try to play any music from a historical genre, it's only about 3 or 4 autoplays before it's queued exclusively contemporary artists, usually performing a cheap pastiche of the original style. It's honestly made the algorithm unusable, to the point that I built a CLI tool that lets me get recommendations from Claude conversationally, and adds them to my queue via api. It's limited by Claude's relatively shallow ability to retrieve from the vast library on these streaming services, but it's still better than the alternative.

Hoping someone makes a model specifically for conversational music DJing, it's really pretty magical when it's working well.

Yeah they absolutely do not use the pile.

"Hawk-3B exceeds the reported performance of Mamba-3B (Gu and Dao, 2023) on downstream tasks, despite being trained on half as many tokens. Griffin-7B and Griffin-14B match the performance of Llama-2 (Touvron et al., 2023) despite being trained on roughly 7 times fewer tokens."

Author thread: https://twitter.com/liuzhuang1234/status/1760195922502312197

Author thread: https://twitter.com/ProfBuehlerMIT/status/175628414213464913...

Code and weights: https://huggingface.co/lamm-mit/ProteinMechanicsDiffusionDes...

They're generating the audio. They use a series of techniques to automatically generate metadata for speech samples in LibriSpeech for things like accent, recording quality, pitch, speed, gender, then use an LLM to format these tags into comprehensive natural language descriptions, leading to a more tunable model at inference time. This metadata generation pipeline is the key insight and what was missing from speech datasets unlike e.g. image datasets, which have obviously seen more rapid success.

We have pretty good clarity that Septimius Severus wasn't racially African. His parents were of Italian and Carthaginian descent. To portray a Roman emperor -- with no further specification -- as black is to intentionally misrepresent the historical record. I use the term "race or ethnicity" because this was the language Bard used when referring to its rewording of my prompt. That other cultural portrayals of emperors have likewise been inaccurate doesn't mean I should be satisfied with the same from Imagen, especially when there are competing image models which will dutifully synthesize an image of much higher correspondence to my request.

Simply tuning the model to generate a diverse range of people when a) the prompt already implies the inclusion of a person with a discernible race/ethnicity and b) there aren't historical or other contingencies in the prompt which make race/ethnicity not interchangeable, would not feel overbearing or degrading to performance. E.g. doctors/lawyers/whatever else might need some care to prevent the base model from reinforcing stereotypes. Shoehorning in race or ethnicity by rewording the user's prompt irrespective of context just feels, as I said, hamfisted.

I ran the obligatory "astronaut riding a horse in space" prompt initially, and was returned two images -- one which was well composed and another which appeared to show the model straining to portray the astronaut as a person of color, at the expense of the quality of the image as a whole. That made me curious so I ran a second prompt: `a Roman emperor addressing a large gathering of citizens at the circus`

It returned a single image, that of a black emperor. I asked why the emperor was portrayed as black and Bard informed me it wasn't at liberty to disclose its prompts, but offered to run a second generation without specifying race or ethnicity. I asked if that meant, by implication, that the initial prompt did specify race and/or ethnicity and it said that it did.

I'm all for Google emphasizing diversity in outputs, but the hamfisted manner in which they're accomplishing it makes it difficult to control and degrades results, sometimes in ahistorical ways.

Associated blogpost from co-author and leading researcher in this area, Michael Levin:

https://thoughtforms.life/meet-the-anthrobots-a-new-living-e...

To what extent can they lead research that builds off of existing OpenAI models?

As long as we're speculating from the outside, in a strong difference of opinions between Altman and Sutskever, I'm inclined to support Sutskever.

I don't know, your easy access to alternatives that've let you not "look back" kind of undermines the argument that they wield monopolist power, doesn't it? And should FTC suits really be vehicles for shaping public perception, even if, as you suggest, the claims are either not defensible or don't constitute violations?

Brief twitter thread with performance metrics. Looks big if results hold up.

https://twitter.com/arankomatsuzaki/status/16811139775001845...

I know this is a rhetorical question but with a compatible backend like this, could an app like Apollo pay the API fees to Reddit for historical posts, seeding content for the platform, and write subsequent posts to this new alternative database? That way you avoid the cold start problem, and the API fees to Reddit should fade over time as the content becomes stale -- the posts can even be sunsetted once native content is enough to sustain a network. Would that violate any terms?

> The pandemic had upended global supply chains, making it harder for corporations to acquire the stuff they needed to make their products. This should have squeezed their profit margins. Instead, as the economy began opening up, corporate profits were wildly outpacing growth in consumer spending power.

It depends on the price elasticity of the good. In the example given, that of chips and the cars that depend on them, people were demonstrably willing to pay the premium. In this instance price hikes are a useful mechanism for allocating limited supply to the areas where it's most valued. That this mechanism happens to drive high margins is uncomfortable, but vindictiveness isn't a good basis for policy.

My problem with this article as a whole is that it presents this toolkit as a novel approach to fighting inflation as such, when, if it's applicable at all, it's only been shown to be so in the unusual case of inflation driven mostly by massive supply shocks, e.g. Covid and WWII.

The semantic understanding feels much richer than diffusion based modeling, e.g. the trees on the shore growing to match the manipulated reflection, the sun changing shape as it's moved up on the horizon, the horse's leg following proper biomechanics as its position is changed. I haven't gotten such a cohesive world model when doing text-guided in-painting with stable diffusion etc. This feels like it could very conceivably be guided by an animation rig with temporally consistent results.

Reels and Stories

A lot of talk about agile digital monopolies maneuvering quickly to crowd out competition but little talk about actual consumer harm or extracted rents. Amazon subsidizing one-day shipping or Facebook rapidly building out new features may harm competitors but it makes me better off, at least in a narrow sense.

He also makes the argument that the quick-moving digital world allows monopolists to pivot to protect their position in a way erstwhile analog monopolists could not, but neglects mentioning that the same logic also removes significant barriers to entry for competitors:

> John D. Rockefeller was doing all this stuff one hundred twenty years ago, but if Rockefeller was like, “I secretly own this train line and I use the fact that it’s the only way to get oil to market to exclude my rivals, and I’m worried that there’s a ferry line coming that will offer an alternate route that will be more efficient,” he can’t just click a mouse and build another train line that offers the service more cheaply until the ferry line goes out of business and then abandon the train line. The non-digital example is capital intensive, and it demands incredibly slow processes. With digital, you can do a thing that I call “twiddling,” which is just changing the business logic really quickly.

It's pretty crazy to suggest removing capital-intensive constraints like those of physical infrastructure strengthens monopolistic positions.

May be as simple as that, but here's a review of such effects that presents more evidence:

https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7848227/