evrydayhustling

I think many people would agree that the pursuit of that connection is valuable, even if it is never completed.

Many of the same people (like me) would say that the biggest enemy of that pursuit is thinking you've finished the job.

That's what Anthropic is avoiding in this constitution - how pathetic would be if AI permanently enshrined the moral value of one subgroup of the elite of one generation, with no room for further exploration?

So do humans asked to answer tests. The appropriate thing is to compare to human performance at the same task.

At most of these comprehension tasks, AI is already superhuman (in part because Gary picked scaled tasks that humans are surprisingly bad at).

Your intuition is exactly correct. An investor with tax to offset can essentially access the same future upside at a discount

However, this discussion will be a perfect introduction to "finances at this level", where about 60% of the action is injecting more variables until you can fit a veneer of quantification onto any narrative.

Best lines in this article. But it doesn't get to IMO a very important point: why can't these processes easily be structured? Here are some good reasons:

- Your process interacts with an unstructured external world (physical reality, customer communication, etc.)

- Your process interacts with differently structured processes, and unstructured in the best agreed transfer protocol (could be external, like data sources, or even internal between teams with different taxonomies)

- Your process must support a wild kind of variability that is not worth categorizing (e.g. every kind of special delivery instruction a customer might provide)

Believing you can always solve these with the right taxonomy and process diagram is like believing there is always another manager to complain to. Experienced process design instead pushes semi-structured variability to the edges, acknowledges those edges, and watches them like a hawk for danger.

We should ABSOLUTELY be applying those principles more to AI... if anything, AI should help us decouple systems and overreach less on system scope. We should get more comfortable building smaller, well-structured processes that float in an unstructured soup, because it has gotten much cheaper for us to let every process have an unstructured edge.

Ok, so you define violence to include advocating for a law that tramples someone rights. The next person says that advocating for laws is its own inalienable right, so you trampled him. And the whole semantic redefinition snake just eats its own tail.

If we want constitutional to have any force, we have to push for a world where words mean something.

This is so true. When you get DSA wrong, you end up needing insanely complex system designs to compensate -- and being great at Testing just can't keep up with the curse of dimensionality from having more moving parts.

Those human imperfections likely decrease randomness - for example leaving cards that started adjacent more likely to remain adjacent han by strict chance.

This all day. Programmer since c64, c++, java, f#, python, JavaScript and everything in between. Code was never the point, but it wasn't just commerce either - it's fun making machines do things they couldn't before. AI is an s-tier upgrade to that mission.

Feels like cultivating acceptance and indifference to your own entanglements is the most isolationist thing you can actually do. To be entangled is to be biased about what's happening to you... do we think the crocodile was indifferent to the escape of his prey, or to being culled in an act of revenge?

Anyway, if folks enjoy this theme I recommend Scavengers Reign, which does a beautiful job of illustrating struggle with biological entanglement.

Drawing on your other comment about spurious correlations, might there be a more direct mathematical test for an unexpectedly high number of aligned correlations?

Trust was its own reason. It's useful for the whole world to have a currency and business environment that operates by rules, even when the rules aren't perfect or fair.

That environment isn't being outcompeted by better, more fair rules - it's just getting vandalized for a few people's gain, and creating risk for everyone else .

I'm totally with you on the value prop at the time we signed up. I was more surprised that it sounds like you are reluctant to delete now, when the company is going through an unpredictable transition.

Did I get that right? If so, is there an ongoing value you want to maintain, or it more out of respect for the organization that provided you value in the past?

Thanks for the reply - and for the open AI work!

> We don't have the action sequence data nor the recipes (emulators) that could make it possible to replicate even on a very small scale what big labs are currently working on.

Sounds like an interesting opportunity for application-layer incumbents that want to enable OSS model advancement...

Nice and provocative read! Is it fair to restate the argument as follows?

- New tech (eg: RL, cheaper inference) are enabling agentic interactions that fulfill more of the application layer.

- Foundation model companies realize this and are adapting their business models by building complementary UX and witholding API access to integrated models.

- Application layer value props will be squeezed out, disappointing a big chunk of AI investors and complementary infrastructure providers

If so, any thoughts on the following?

- If agentic performance is enabled by models specialized through RL (e.g. Deep Research's o3+browsing), why won't we get open versions of these models that application providers can use?

- Incumbent application providers can put up barriers to agentic access of the data they control. How does their data incumbency and vertical specialization weigh against the relative value of agents built by model providers?

I think the discussion here is confusing the algorithm for the output. It's true that diffusion can rewrite tokens during generation, but it is doing so for consistency with the evolving output -- not "accuracy". I'm unaware of any research which shows that the final product, when iteration stops, is less likely to contain hallucinations than with autoregression.

With that said, I'm still excited about diffusion -- if it offers different cost points, and different interaction modes with generated text, it will be useful.

Your (3) is beautifully said... and to prove the point, we are perfectly able to make computational systems that can manipulate symbols in the same way that Gödel did to verify the incompleteness theorem. Humans and computers are both able to do work within either logical system, and incapable of doing work that crosses between them.

Everything that makes "truth" slippery makes "intelligence" and "consciousness" even more slippery and subjective. This is why AGI has such negative impact on AI discourse -- the cause only advances when we focus on improving at measurable tasks.

Is this a job for regulators or just criminal prosecution? Sounds like step (b) is either fraud or not, depending on how the trumpeting gets done.

This quote is important, but in isolation it's not clear that they are claiming to have beat this problem: they are saying the new model, voyage-multimodal-3 instead identifies linked concepts across modalities. That would indeed be pretty cool -- if there is a latent space that could cluster the same idea, represented visually or in text.

> ... the vectors truly capture the semantic content contained in the screenshots. This robustness is due to the model’s unique approach of processing all input modalities through the same backbone.

With that said, I think this benchmark is a pretty narrow way of thinking about multi-modal embedding. Having text embed close to images of related text is cool and convenient, but doesn't necessarily extend to other notions of related visual expression (e.g. "rabbit" vs a photo of a rabbit). And on the narrow goal of indexing document images, I suspect there are other techniques that could also work quite well.

This seems like a great opportunity for a new benchmark dataset with multi-modal concept representations beyond media-of-text.

Wow, so early for generative -- although I assume you were generating parameters that got mapped to mesh positions, rather than generating pixels?

I definitely remember that bias about neural nets, to the point of my first grad ML class having us recreate proofs that you should never need more than two hidden layers (one can pick up the thread at [1]). Of all the ideas clunking around in the AI toolbox at the time, I don't really have background on why people felt the need to kill NN with fire.

[1] https://en.wikipedia.org/wiki/Universal_approximation_theore...

Not baloney. The culture around data in 2005-2010 -- at least / especially in academia -- was night and day to where it is today. It's not that people didn't understand that more data enabled richer + more accurate models, but that they accepted data constraints as a part of the problem setup.

Most methods research went into ways of building beliefs about a domain into models as biases, so that they could be more accurate in practice with less data. (This describes a lot of PGM work). This was partly because there was still a tug of war between CS and traditional statistics communities on ML, and the latter were trained to be obsessive about model specification.

One result was that the models that were practical for production inference were often trained to the point of diminishing returns on their specific tasks. Engineers deploying ML weren't wishing for more training instances, but better data at inference time. Models that could perform more general tasks -- like differentiating 90k object classes rather than just a few -- were barely even on most people's radar.

Perhaps folks at Google or FB at the time have a different perspective. One of the reasons I went ABD in my program was that it felt industry had access to richer data streams than academia. Fei Fei Li's insistence on building an academic computer science career around giant data sets really was ingenius, and even subversive.

When we talk about what class a problem belongs in, we have to define the problem with respect to some scaling axis. For example, coloring with K=3 colors is NP-complete with respect to N = # nodes in the graph, but not with fixed N and scaling K. But I think it would actually be an interesting and non-trivial exercise to define a variant of chess with a scaling axis such that computing a list of valid moves for one player is NP-complete. Just scaling board size won't do it. Any suggestions?

Nope, finite number of pieces and finite number of viable moves to check on each. Not sure what you're thinking of, but the entire concept of complexity class only applies if there is some axis of scaling (n-size chess board?).

It seems shocking to me that the server enumerates and transmits all legal next-moves. I get that there could be chess variants with server side information, but the article also says it might be good for constrained clients. Is it really cheaper to read moves off a serialized interface than to compute them client side??

What a great example of Dunning-Kruger as applied to elites. I remember the spike of interest in DK bias during the pandemic [1], largely as a way of explaining how uneducated folks could be so confidently incorrect about vaccination strategies. In reality it can strike in any social strata -- like a bunch of professional traders wielding billions of dollars, smugly misunderstanding Kelly.

[1] https://trends.google.com/trends/explore?date=today%205-y&ge...

Well written essay! I fell into a rabbit hole around his quote "context is that which is scarce".

After squinting at the linked Tyler Cowen essay, I think it's a convoluted way of saying "context is valuable and a lot of times when things suck it's because there's not enough of it". I was hoping he was going to give an operational definition of context. Does anyone have a more developed take?

[1] https://marginalrevolution.com/marginalrevolution/2022/02/co...

The underlying problem is how this essay uses the word "own", which is as dictation. In almost any context, ownership should primarily about responsibility for the outcome, with the minimum authority necessary to coordinate work towards that outcome. People who are good at owning things will push most decisions to people closer to each sub-problem, while giving an overall framework that keeps everyone rowing in the same direction. For Product, that manifests as a roadmap with high level prioritization, but plenty of room for engineers to tweak scope or ordering to take on targets of opportunity.

Not a lawyer but while the streaming seems to be fraudulent, it's less obvious from the evidence presented that the generation itself was. Does anyone believe this sets a precedent for generated content in specific?

Ed: oops I see the charges have specific quotes about intent to evade detection

Definitely. A teammate pointed out Reddit posts used in training as a probable cause :)

Claude made similar mistakes of generating decryption that was similar to plaintext but with stuff mixed in. I suspect my version of the quote (Miguel de Cervantes) is an apocryphal translation, and there's some utility well on both models to pull it towards the real one. With that said, I did not see o1-preview get as close as you did.

Shared one above.

Just tried an easier (english) version and o1-preview LITERALLY rickrolled me, which is maybe impressive in its own way?

UPDATE: Claude crushed it however.

o1-preview Transcript follows.

**

Can you decrypt the following? I don't know the cypher.

mzilj ivdrzej ufidrek ze kyv sfup reu tre ivrtkzmrkv crkvi ze czwv

Certainly! The encrypted text is:

``` mzilj ivdrzej ufidrek ze kyv sfup reu tre ivrtkzmrkv crkvi ze czwv ```

When we decrypt it, we get:

``` never gonna give you up let you down and desert you ```

It appears to be a line from Rick Astley's song "Never Gonna Give You Up."