Hacker Newsnew | past | comments | ask | show | jobs | submit | cs702's commentslogin

> The six incidents included models developing ways to ignore “normal constraints”, fabricating and misrepresenting data and hiding mistakes made while conducting tasks, the San Francisco-based company said on Wednesday.

every model does this, it's their jam

Teenagers..cant wait to kick them out of the house...

just wait until the shoe is on the other foot and they are saying "can't wait to put them in the old folks home"

funny how life has two sides like that, we start out needing all the help and we largely end there too


~85% accuracy on MNIST. Sigh.

How does it do on CIFAR-10, or even better, ImageNet?

Interesting research, not sure it's a backprop alternative.

===

EDIT: accuracy on MNIST is not ~90%. It's ~85%.


It might be beneficial while not being optimal on its own.

The obvious example is if it has different behaviour around local minima, it could be an altenate pathway out.

I have often wondered if doing training with radically different aproaches for the first few iterarions would avoid any method specific artifacts before the weights had time to denoise.


Yes, it's interesting research, I agree and even wrote so in my post :-)

But the submission title calls it a "backprop alternative." It is not, at least not yet.


You can probably distribute training more easily too

They state replacing backprop is not their goal.

Their goal is to understand how distributed systems which cannot do backprop (the brain) can still do learning.


Personally I think the dirty secret of the brain is that a lot of things are hard coded. And many things that we need to learn are also hard coded except that some parameters need to be tuned.

If we puke, the brain will not do general aversive learning, it will learn to avoid specifically the last thing eaten, because it instinctively knows about food poisoning.

Imprinting is absolutely fascinating. Some newborn animals will run a very simple pattern detector like looking for a red dot or something and use that to bootstrap their conception of their parent.

For fully general learning I have a hunch that it can be done using local history plus a semi-global reward scalar (global neurotransmittor levels).


regardless if the intelligence in the brain is hardcoded or not, to the extent it is, this information must have been compressed in the genome, which runs counter to almost all observations: a child doesn't remember the experience of their ancestors, for example. The only sense in which we do carry mental state without relearning is emotions, instincts, reflexes (some neuronal pathways that connect the eye to the middle ear), hormonal driven behavior (fear adrenalin).

For another, there are about 200k promotor regions (including non-coding) in the human genome.

A promotor region might have say 6 to 15 bits of information.

Can you compress 2025 or even 2024 era LLM intelligence into 3 megabit = ~400 kB ? I think not. I think a lot of compression is still possible, but 400 kB?

So I think we can box up the idea of "dirty secrets of the braing: not learning but hard coding". There is a lot of hard coding in biology, but brains are evolved specifically to enable learning within the individual lifetime instead of only learning by natural selection.

I also don't buy the following argument:

> If we puke, the brain will not do general aversive learning, it will learn to avoid specifically the last thing eaten, because it instinctively knows about food poisoning.

Each time it happens that I end up puking, I do feel aversion and try to avoid puking at all, sometimes I succeed but sometimes is just puke. There must be fundamental puke reflexes (which one fails to avoid) and avertable puke reflexes.


You are thinking on the wrong level. Of course we don't have an encyclopedic knowledge of the world encoded into our genome. It's like we have certain structures of the world hard coded, and they may use different learning algorithms.

We instinctively know that there are other intelligent beings, and we have the the machinery to model them. We are born with the capacity for language. It must still be learned, the specific words aren't hardcoded, but the concept of language is. We are born with the capacity to store and replay memories. The brain knows some aspects of how the world is supposed to look like visually, and if it doesn't it will try to correct that. People who used optics to see the world upside down have found that after a brief time their brain learned to flip the world right side up.


> Can you compress 2025 or even 2024 era LLM intelligence into 3 megabit = ~400 kB ? I think not. I think a lot of compression is still possible, but 400 kB?

There are a few extra levels of interpretation (like protein synthesis) that are more like a transpiler than compression (imo), over a 4-base language that is read in a sliding window and is affected by surrounding conditions, so the same "token" sequence may produce different things depending on external factors. Some biologists I used to collaborate with talked about 7 layers to this process, I have only described one level here


From an information theory perspective it does not matter how many levels of interpretation are in between, that hardcoded information must pass the genome. for example we can discuss "what about protein synthesis", for example binding affinities, folding helpers etc. they in turn were encoded genetically as well.

> There are a few extra levels of interpretation (like protein synthesis) that are more like a transpiler than compression (imo), over a 4-base language that is read in a sliding window and is affected by surrounding conditions, so the same "token" sequence may produce different things depending on external factors. Some biologists I used to collaborate with talked about 7 layers to this process, I have only described one level here

so the same "token" sequence may produce different things depending on external factors.

yes, non-hereditary learning depends on external factors, thank you for paraphrasing me while shifting attention.

the multi-scale nature (transpilers etc.) doesn't change the theorems in probability and information theory which seriously constrain the maximum amount of information a message can store.

The whole point of a brain is that it is an organ dedicated to storing, retrieving and timely utilising information one can't afford to store in a genome.


I'm not convinced information theory is the right avenue for such non-deterministic, environmentally affected, open ended systems. Protein synthesis is defined by far more than the DNA, which is always being processed, most of which is junk and being recycled. The layers in between are more than interpretation of the prior layers, the data grows at each level to incorporate more sources under your information theoretic formulation.

I was never paraphrasing you, but thank you for attempting rhetorical antics?


Information theory is the right avenue but it's too often used to imply things it says nothing about. From (1)

1. A determinist transformation can't create more information than its inputs

you can't derive (2)

2. Later stages don't contain any new information.

without restraining your model to a monotonic chain of transformations that can't tap into the information content of the environment.


> I'm not convinced information theory...

I'm not convinced my molecules follow the laws of thermodynamics.


"Backprop alternative" was the OP's headline.

I think my initial assessment is correct: interesting research, but not an alternative, at least not yet.


Ah. I missed that. Yeah. "Backprop alternative" is the wrong framing.

Thanks. The exaggerated framing is the only reason why I posted my comment :-)

Their image classification benchmarks include both: https://pub.sakana.ai/pc-alm/assets/figures/benchmark_accura...

~74% on CIFAR-10. Still a far cry from backprop.

I didn't see ImageNet. TinyImageNet is something else.


There might be only one paper that's trained on full Imagenet using methods like these.

Training a Predictive Coding Network on ImageNet using Equilibrium Propagation Tugdual Kerjan, Rasmus Høier, Benjamin Scellier https://arxiv.org/abs/2606.03584

It's quite an undertaking.


It's really sad that they are only a few years away from making backpropagation completely obsolete.

"There are two things that are important in politics. The first is money and I can't remember what the second one is." - Mark Hanna

Mark Hanna was a political consultant in the late 1800's, in the US.

Source: https://quote.org/quote/there-are-two-things-that-are-import...


To me, it looks like the leading labs are investing more and more to improve their frontier models, and are being forced to charge less and less for them due to competitive pressure.

Maybe I'm wrong, but "reasoning as a service" is looking more and more like a... commodity.


Reminds me of the Black Mirror episode in which a blue animated bear named Waldo nearly wins a by-election:

https://en.wikipedia.org/wiki/The_Waldo_Moment

Reality these days sometimes feels as strange as fiction.


Yes!


I mean that one came true in 2016.


That is my reaction too. It looks like great work!

Valuable not only for inference, but for training too (think proprietary datasets).

I would add, a single individual did this.

One person can make a difference :-)


The OP talks about the drought extensively. Quoting:

> there is quite a lot of compelling evidence that period of LBAC [late bronze age collapse], especially the 1190s, was unusually dry in the Eastern Mediterranean, which would have caused reduced agricultural output (crop failures). Interestingly, this would be most immediately impactful in areas engaged primarily in rainfall agriculture (Greece, Anatolia, the Levant) and less impactful in areas engaged more in irrigation agriculture (Egypt, Mesopotamia).³ And, oh look, the areas where LBAC was more severe are in the rainfall zone and the areas where it was less severe are in the irrigation zone.


One possibility I've wondered about is the emergence of a new crop pathogen. This might be addressed by looking at DNA of modern crop pathogens, and possibly looking if there was a change in the crops being grown before/after the LBAC.


It's not hard to imagine that will be able to do as good a job as a human accountant in the not too distant future.

It's also not hard to imagine tax authorities using AI to audit everyone's tax returns every year.

We sure live in interesting times.


The real test to see if AI is just a rich person thing will be to see how the tax authorities treat it, even for more complex returns.

They can save humans for the really complex edge case stuff but at the end of the day, the tax code is just checkboxes and input forms that get boiled down into Integers, Floats/Doubles and enumerated choices with some Strings for deductions


I've submitted my German taxes this year using a mix of Claude 4.6 and Claude 4.7, with lots of manual checking. The German Finanzamt granted most of the things I listed in the tax return (they send you an official letter by post) -- I did have to appeal for one of the items though (again using Claude, this time 4.8 ).

The most important thing I've found is to ask Claude to thoroughly audit the reply (to find all hallucinations). I usually ask it to give me an enumerated list of all facts and all legal cases quoted, and then I give it to a new instance to carefully validate each one.

Newer models are getting much better at not hallucinating German case law though :)


Have you considered using using two completely different models and comparing their output in order to catch hallucinations?


Bookkeeping is not tax preparation, just FYI. The reports generated by an accounting system are used on tax returns for companies, but they’re distinctly different things.


Also, in case you didn't already know, I saw a headline announcing that the sky is usually blue.


Bending Spoons is a company that acquires SaaS companies/products that are not growing or losing users but have a well-known brand and customers who stick around.

The execs at Bending Spoon buy these SaaS services on the cheap, cut costs, jack up prices, and milk remaining users for as much cash as possible for as long as possible.

Rinse and repeat. The goal is to generate the highest possible rate of return on invested capital in a law-abiding manner.


And the article tries to spin this positively:

> After the acquisition, Bending Spoons is anything but a passive owner, making changes to the products’ user experience and features, as well as to the underlying tech; monetization strategy, including pricing; and team organization, including headcount.

> While this focus on efficiency and revenue overlaps with private equity strategies, Bending Spoons claims a key difference: It “aims to hold forever, and has never sold an acquired business.” It is building a live portfolio, not presiding over a tech graveyard.

That last line has me wondering who wrote this.


Renowned author C. H. Atgpt


> building a live portfolio

The portfolio grows -- "live", and faster than each of the businesses within it shrink. As they all do, by Bending Spoons' strategic design.


I don’t feel like the article was sortballing the company. They brought up things like the WeTransfer founder criticizing Bending Spoons’ decisions.

As for my opinion on the company, I don’t really see anything particularly negative about it. I think the fact that they’ve never sold an acquired business is a rather admirable trait.

In a way, they’re doing something that may not have been possible without this style of intervention, which is to keep companies/products that would have otherwise disappeared viable.

For a company like Evernote it wouldn’t be better for their customers if the company liquidated. There are worse things that can happen to your service provider of choice than price increases or worse customer support.


People are framing this like they're creating sustainable businesses, but if you look into the details, what they're consistently doing is stagnating on any kind of feature development, making the apps and sites more difficult to use and have more nags, and they're increasing prices, sometimes by 10x or 100x. When I look for a company that I think I would admire, I'm looking for customers that are satisfied and recommend the product to their friends.

Charging $20,000 for a note-taking app subscription is not that.

https://news.ycombinator.com/item?id=48849810


I don't know about their other software, but i'm a (paying) user of komoot. Komoot development has stagnated until bending spoons took over ~1y ago, see the recent https://www.komoot.com/product-updates

maybe it's case by case basis; i'm not an evernote user since the botched rewrite probably 10y ago at this point


I certainly don’t find any of that positive, either, but sometimes what a lot of these companies need to survive is to increase prices and only worry about the feelings of the customers who find those higher prices to be worth it.

The $20,000 price plan wasn’t a real price, that was just a not so gentle nudge to move to a different offering. Maybe it feels bad but that plan effectively doesn’t exist anymore. Things change.

It’s got fewer features for the dollar, but if the previous company was not sustainable in the first place, it is what it is.

A company raising prices or cutting service quality is only a problem if they’re in a monopoly situation with no other market alternatives. None of the companies Bending Spoons has acquired are in that position. Many of them are far from being the market leaders.

The point is that Bending Spoons isn’t buying companies and saddling them with unsustainable debt like they’re Toys R Us. They’re buying companies that need drastic operating change and implementing that change so that they can exist in perpetuity.


> Things change.

This is true, but there are choices you can make in life to really minimize the impact. I've been using TiddlyWiki for more than 20 years and it always Just Works. I picked it precisely because I value endurance in the software I choose. I know that's not a fad right now, but folks just don't have to subject themselves to this standard of treatment.

That aside, my objection is the use of shady tactics to achieve that goal (constant nags and popups, massive price increases for reduced service, rejection of previously "lifetime" memberships, etc.), at the expense of the customer. Swaddling that in a blanket of "it's sustainable" makes me feel only a tiny bit better about it. To make an extreme comparison: fraud is also sustainable; I guess I'm saying sustainability is not an inherent good. If (hypothetically) every Evernote customer would be better off if they were using Joplin, keeping Evernote around would be a bug, not a feature. I don't think this is actually true, might it might be close.


I think that every time an option leaves a market it’s a detriment to competition. It doesn’t really matter if that option was bad or that I won’t personally choose it.

I’m certainly very much against lifetime subscriptions losing promised features and things like that.


I will add to my original reply, if I drink their company kool aid on their company website they pretty specifically list out a number of improvements they’ve made to their product portfolio.

They have claims like making Evernote sync faster, fixing stuck transfers on WeTransfer, offering a free organizer EventBrite account for the first time since 2005. These seem like pretty tangible claims.

Perhaps they are trying to combat this exact negative image that they’re just there to suck out value.

Maybe they’re lying about their accomplishments, I really don’t know. I don’t use any of their products.


Talking to real users tells a different story: https://news.ycombinator.com/item?id=48847994


Sure, one user’s anecdote. You can find negative reviews for any product.

Evernote on iOS currently has a 4.4 star rating.

For comparison, the Obsidian iOS app has a 4.5 star rating.


Yep, I ended up looking into this quite a bit more, and now agree with you. Evernote ratings were historically lower before the Bending Spoons acquisition. There are shady things you can do with ratings like ask whether they like they app before redirecting them to the rate it, it does seem like I was over-indexing on some anecdotes (I also worked with some ex-Evernoters, so I was coming in a bit biased). Thanks for pointing out the trend!


There are four stages to any successful companies lifecycle and Bending Spoons's model is to maximize what they can get in the final stage of decline.

There's nothing wrong with that, but if you're a user of one of these services you might take it as a hint to find an alternative.


> There are four stages to any successful companies lifecycle

I usually say in interviews that my preferred time to join a company is at the end of stage 1 (start up) and the start of phase 2 (organizing).

Nothing makes me happier than to be told "Hey, we got this up and running and it's a mess. Now we need someone to turn this into a system that is easy to modify and maintain."


Yeah, gotta call out the "There's nothing wrong with that".


Can a business not just reach a steady state after maximizing its market?


It makes no economic sense for AOL to stay in a steady state after having been thoroughly disrupted by newer products.


“ There's nothing wrong with that”

Debatable


There's skill in being able to manage a declining or non-growth business in a way that still pleases your consumer base (and therefore reduces your attrition rate). Not everyone does it well.


What you said is not synonymous with “maximize what they can get in the final stage of decline.”.


There's nothing wrong with it per se. Plenty of great products were ruined because management refused to accept that it wasn't in a growth market anymore and should be run for minimising customer losses, not gaining substantial new ones. That, in turn, means laying off a lot of the design, engineering and sales talent that was necessary for the previous configuration.

You can also be a bastard and jack up prices while cutting e.g. customer service. (Though absent new major revisions, service costs should go down.) But I'd argue we need, in tech, more of this strategy of calm wind-down than the everything-must-be-growth mindset.


Is this practice good for the workers in the organisation?


Definitely, no.


yes, they are trying to gouge me on Evernote that no longer works, that i tried to unsubscribe from

here is a solid article from this week's Economist (that mentions another real jewel of a company):

https://www.economist.com/business/2026/07/01/can-bending-sp...


Similar story here. They took my ~$100/yr Harvest time-tracking Solo plan, increased the price by 2.5x for a more restricted plan than I had... or I could get back the plan I had for $20,000/year.

So I downloaded my data, and had Claude vibecode a fully-featured clone in a single evening. Even if I was paying Anthropic API rates, it cost me less than a single year of my Solo plan.


Was also on Harvest when news broke they had bought them here on HN. A lot of the same comments. I thought, "Well, maybe this is hyperbole, let's wait it out." About a month after they were acquired, same thing. Price of my plan went up almost by double.

So if anybody is reading this? They absolutely will gouge you. All the stories you've read are all true. Take some advice and get out while you can.


Double? Rookie numbers.

We went from $1k to $21k per year.

A few emails later and we have a hefty discount, but we will still move away.


Tip: It's straightforward to migrate from Evernote to Apple Notes. https://www.macrumors.com/how-to/migrate-from-evernote-to-ap...


eh, i backed up in a few places a bit ago. the actual concern is BS charged me when they shouldn't have (5x the former price, annual), won't refund, and turned off the account anyway. and PayPal seems to have an open marriage with PCI-DSS/SOC2 right now


I've moved to Joplin and am pretty happy with it. Was easy to self host on my WebDAV.

I believe it imports Evernote data too.


I had a vendor acquired by one of these types of outfits.

I looked through their assets and it clicked: “this is where software goes to die”


> this is where software goes to die

The ones that IBM passed up on, yes.


That avoided the Apache Foundation.


That's a short term business model if I have ever seen one.

"customers who stick around." is anthesis to mid- to long-term customer loyalty when you do "jack up prices, and milk remaining users for as much cash as possible"


Think of it as a perpetual bond with declining coupon payments.

Customer "inertia" or "lock-in" might be better terms to describe what the company is looking for in an acquisition.

Their ideal customer may well be someone who's forgotten they have a subscription on credit card auto-pay.


> a perpetual bond with declining coupon payments

Most things with royalties (oil fields, songs) work like this.


Yes, agree.


Add to this that they make it really, really hard to unsubscribe. I think there's been some legal crackdowns, but for a time, they could make it literally impossible.


correct, they have made it impossible, charged my 2002-era PayPal account when i said, "i want to leave, don't"


This article is like an advertisement. Here's how they spin it:

> Speaking to TechCrunch, co-founder and chief product officer Matteo Danieli said some of the scrutiny was due to the fact that products such as Evernote were genuinely loved by their users. But he said that despite all the changes, customer retention has been “remarkably stable.”

Ah yes. In other news, the prison population size is also remarkably stable.


You're thinking too narrowly. Buying a cow is a short term investment because cows don't live very long.

And yet dairy farms can last for centuries.


They reproduce.

I haven't seen Evernote pro- nor reproduce anything


I'm now imagining two little SaaS getting together and making another small little SaaS.


> cut costs, jack up prices, and milk remaining users for as much cash as possible for as long as possible.

Don't forget "slash the workforce, ensuring that the product will get worse over time".


> Don't forget "slash the workforce, ensuring that the product will get worse over time"

Not commenting on Bending Spoons. But in general, a company built to grow is overprovisioned for one being put into maintenance mode. If you're growing, sure, let the designers change the UI every release. If you're trying not to lose customers, don't do that. Which means you don't need a crack team of in-house designers.


I agree that a company in growth mode needs more employees than one in maintenance mode. But wouldn't the owners already have cut unnecessary employees before selling out to PE or similar?


> wouldn't the owners already have cut unnecessary employees before selling out to PE or similar?

Usually not. It's emotionally difficult. And knowing what you need and don't need to cut (versus transition or aggregate with your conglomerate's administrative layer) is its own expertise. If you had that, you wouldn't need Bending Spoons or whomever.


That’s what “cut costs” means for SaaS, firing people.


Yes, it is part of cutting costs, but there are other ways to cut costs ("synergies" from merging back office functions) that don't necessarily affect the product.


Consolidating stagnant or dying SaaS offerings makes sense, but it'd be nice if there were a version of this that's a better steward of the companies.


If that was good business then presumably the brand could have done it at some point during their long slow decline?


Arguably, they're a better choice for customers than a shutdown.

I mean, they're at least keeping the service alive for as long as possible.


There’s no choice here, and often the companies are profitable, but if there is any stickiness to the product the customer gets the privilege of having a company they built trust with turn around and betray them with massively increased fees.


I'd argue it's worse for consumers, by keeping them alive it staves off competition, and leeches cash by increasing subscription prices or locking once free feature behind paywalls.


So like Computer Associates?


So like Broadcom?


So like Delphi?


> The execs at Bending Spoon buy these SaaS services on the cheap, cut costs, jack up prices, and milk remaining users for as much cash as possible for as long as possible.

If they are such stable long term SaaS businesses who aren’t losing customers, why are they selling to bending spoons?


because there's no joy in managing a declining company, especially when you made it grow in the past, and probably get enough money from the deal that you don't need to care anymore.


So then we prefer the scenario where the employees are just forced to manage a declining company? Or we prefer it folding?


Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: