deepsquirrelnet

It's hard to say without knowing their architecture, but I'd guess something like block attention. You can process the prompt separately from the classifications into a latent space and then do some kind of late interaction with the encodings from the classifications.

There are plenty of other ways to do zero shot classification that would result in more "token usage" (really just having to reprocess everything for each class), but the pricing and the way they describe it narrows it down somewhat.

I'm having a really hard time wrapping my head around why Jev is getting so much hype. It feels manufactured to me. I don't think they've proven a significant market for their product, and there's no independent benchmarks that prove anything. To me that's doesn't pass the smell test.

But if they had to show how well their product worked they might give away the whole game... because they'd have to compare their "noul" class against an NLI benchmark for instance, and possibly show they're losing to cross encoders and give away the fact that they are just rebranding NLI. Or rerankers (choice) or zero-shot classifiers.

I'm not sure anybody but people inside the company know if the model itself is a contribution. There's no publication and no architectural details. There's no benchmarks or comparisons published. You can do all of the things they claim with an LLM, not that I think that's what they did.

Likely they have some encoder (eg ModernBERT) trained to do late interaction or latent states along the lines of ColBERT, Perceiver IO or poly-encoders.

No. Frontier AI companies need the government to shield them from becoming commodity. That's why they're begging for regulation.

On the other side, the current administration also has powerful backers whose businesses are complemented by AI and would benefit from it becoming a commodity. Think about who Vance is speaking on behalf of.

So the AI companies need to use fear to drive public demand for their regulation efforts. IMO there's little to no honest discourse and a lot of attempts to manipulate public opinion for objectives that we have little stake in. I personally don't trust any of the involved parties to make decisions for the public good.

Wow, how could this have happened right before the election? Surely this will not be used as a pretext for anything.

The problem is that at the business level and at least for a large part of the US government, we need to assume that there are no 'good' actors.

Anthropic needs regulation in order to prevent AI from becoming a commodity. This of course does not benefit all tech businesses equally, especially those that are not currently at the AI frontier. So when JD Vance talks about AI, he talks using the mouth of Peter Thiel who may not see benefit from the same policy as Altman or Amodei.

The rest is just public support posturing and most of that is bullshit meant to distract from the high rollers game of winners and losers. The philosophy is money and power, who gets it and who doesn't. Us normies aren't really participants in the game, except where we are being manipulated into cheering for one side or another, and with little stake in the outcomes (although selfishly, I'd be pissed if I didn't have open models to tinker with).

Open source is fundamentally a vehicle for commoditization. This is great if your business is not AI and your business is instead something like GPU hardware or some product that uses AI. But it means that eventually, selling AI is not going to be the money maker.

OSI proliferated open source on a business strategy called "commoditizing your complements". These big companies don't do it out of benevolence. It was pitched to them in a way that FSF did not (which was more about morals and ethics, something business care little about), and it caught on. And the software business became about ads, consulting and cloud services instead.

It's also clearly the wrong question for this decade. Private / public sector cooperation and all of the barriers between them that have been degraded and chipped away has left us with a very broken economic system that looks less and less like capitalism every day.

They wouldn't provide results. You'd only get conclusions that were made from the beginning.

> Republican strategists are expecting over 150 million to be spent on this race

Imagine the assurances that are made to convince donors it's worth it. Who exactly bears the cost of unpopularity?...

People called Don and irrational hate of windmills. Name a more iconic duo...

Around 2015, my boss got permission from corporate to take me and a few of my coworkers to a leadership conference for a day. It was setup at a church and had religious theming around it, and a lot of "leadership values" were communicated. It was the usual fare of responsibility, accountability, truth and trust you'd expect, mixed with the religious overtones you'd also expect. But what I didn't expect is how trite this fictional narrative would prove to be.

I think about that a lot now, and how my religious upbringing was bombarded with this messaging (in a good way). And how much of the same world that previously claimed espouse it has complete turned their backs on it in such a short time.

Is it dead or did it ever exist at all? At least we don't seem to pretend to have higher values anymore. If we even have leaders at all, they are not measured by the yardsticks we used to make.

That's a good point and conventionally if benchmarks aren't run as "one shot", it is denoted as "benchmark@K". Inference time scaling has historically shown improvement.

Generally though, many of these fairness complaints do go away if there is "3rd party testing". Right now, companies reporting their own benchmarks has all the problems that 3rd party testing resolves in many other industries.

If you look at papers on benchmarks, they're usually created to expose gaps in how models are trained. It should be no surprise that models get better on them over time, because you can't get better at what you don't measure.

Cherry picking the benchmarks you present is where the falsehoods lie.

Praise FSM. Without circular investments we'd all be broke.

China is doing open models better. I think there's little going in Hagueseth's head other than the usual reactionary nationalism reflex.

I think it's hilarious. LinkedIn is rushing to de-legitimize themselves so hard that they're inventing a new market for someone else to step into. Apparently indeed doesn't want to take it... not sure what's going on there.

> You've used 92% of your Fable 5 limit · resets Jul 12, 12pm

So generous.

Turns out it's easier to make conspiracies than effective policy. Who knew?

I think this is only accurate when no external ideas are used, but I'd like to suggest that nearly all new discovery is built on a combination of old ideas and LLMs are really good at the latter.

If you bring something new to the table, then in my experience, AIs are really good at helping you ground it old ideas. If you want to set it and forget it, then you will get the mean. If you want to do something new, in my experience, they are enablers and not blockers.

Bingo. Value is the operative word here. Money and privacy are valuable too, and I'm assuming that there must be some pitch deck somewhere that is presumably good and selling city councils on this. Where is it? What is the value we're supposed to get from this?

Can anybody find trustworthy stats that these actually reduce crime? All I see are occasional anecdotes about how they were used to find one person one time.

Skeptical me seriously doubts this is an effective solution for crime. But maybe that's because this country has a history of being willing to do a million expensive and privacy violating things, and only if it's a punitive measure.

> Because I love swiping, but all my problems with it come from the fact that the QWERTY layout is far from ideal for it. I am 100% willing to learn a new layout if anyone will develop an optimal one for English so that swiping has a 99.9% accuracy rate instead of what currently feels more like 90% or 95%.

90-95% is a very good estimate! That's about what we measure on our test set. I have good news for you, and we will have a blog post about it soon. Because of how our models are built, we are able to optimize for detection accuracy directly by constructing synthetic swipes on each layout for ~50k words, and then testing them through the model. We tested around 800,000 layouts this way.

The biggest issue with QWERTY is that there are far too many words that swipe colinear or obtuse angle letter trigrams. These are both hard to detect and frustrating for swipe users, because you can't clearly indicate the letters you're gesturing. Neural swipe models (at least ours) look for indicators in the gesture pattern that suggests a user was targeting a specific letter, rather than trying to match a gesture shape like algorithmic detection does.

The shape of the keyboard can significantly improve the way the gestures are formed so that there is better indication of letters. The model can still respond to dwell times because unlike shape matching it uses the temporal information. But dwell interrupts flow, and in my opinion should be minimized in swipe layouts.

I'd been working with language models for several years before LLMs were a solution to this kind of problem. These are some ideas "off the top of my head" about how you can do classification in various ways. There's really a lot of ways to tackle it now, and a lot of trade-offs you can learn by experimenting with them.

There's even more options still, especially if you go further back toward more traditional methods. Static word vectors like GloVe or fasttext (optionally more modern equivalents like WordLlama or Model2Vec). Then there's sklearn-style stuff too. Those can be really small/fast but have more accuracy-level tradeoffs.

If you want to go deeper on language models, try these project ideas:

- Zero-shot encoders like tasksource or GliNER

- Natural language inference: https://huggingface.co/blog/dleemiller/nli-xenc-ways-to-use

- GRPO training

- GEPA prompt tuning Qwen 0.6B (or GEPA, then GRPO)

- Use an embedding model and train a classifier (MLP, logistic, svm)

- Use a larger LLM to generate a synthetic dataset (beware of lack of diversity, mine "seed text" from real sources first)

- Synthetically generate "hard examples" where more than one category may be valid and DPO tune your preferred responses

The first step to regulatory capture is getting yourself regulated...

> Unions exist to benefit the median and bring up the floor, but it stifles competition among those who really do desire to be at the top. And in doing so while it brings up the floor, it also brings down the ceiling because people who would normally be motivated enough to move up would not have much incentive to do so anymore.

I think people tend to fixate on the worker-to-worker differences inside of unions. Yes, that is the most visible part of a union when in place, and at least in the US has valid arguments about meritocracy.

What is missed when limiting the scope to just that is the population-level abuses of workers that no amount of meritocracy will fix. When corporations engage in collusion against workers (now common and nearly unpunished in the US) the top-level wages are suppressed industry wide.

The whole pay band alignment that comes out of that undermines the meritocracy argument, and doesn't even begin to address the wage-fixing that has gone almost unchecked in tech for decades[1,2]. As a merited employee, you might have more options to where you can go, but it won't protect you from predatory hiring/layoff cycles and it certainly won't guarantee that you'll receive a truly competitive wage.

On paper, meritocracy sounds great. I have worked many places in tech and never once observed it, personally. Best case, if you have warmed a seat for enough years, then you advance that way. Worst, your employer knows they can just take advantage of you because you're willing to work without a dangling carrot.

As before, either the government frees itself from corruption and enacts justice or unions will come back. That is point we are at.

[1] https://www.npr.org/sections/alltechconsidered/2015/01/16/37...

[2] https://conversableeconomist.com/2025/10/31/the-silicon-vall...

Baked into that is a presumption of justice, which is becoming comically out of touch to the point where that overused phrase could be a meme.

That’s even a bit optimistic. We can’t agree that all people with full time jobs should be able to afford the basics for their own survival.

Flock is only one company. Someone in my town smashed one from a different company and was treated like a hero in Facebook comments. It’s not mentioned in the article.

Not that I’d want to work there given what they do, but every time I’ve been contacted by a recruiter there, it seems like it’s within a month of a mass layoff they’ve had… which is maybe just because they seem to have mass layoffs every quarter now.

They also seem to have adopted a no-remote hire policy and are in an extreme high CoL location. It’s a truly awful mix for trying to attract outside talent. I don’t know why they even bother.