Closi

It's not particularly interesting if Deep Mind comes to the same (correct) conclusion on a single problem as o3 but costs more. You could ask gpt 2.5 and gpt4 what 1+1= and would get the same response with gpt 4 costing more, but this doesn't tell us much about model capability or value.

It would be more interesting to know if it can handle problems that o3 can't do, or if it is 'correct' more often than o3 pro on these sort of problems.

i.e. if o3 is correct 90% of the time, but deep mind is correct 91% of the time on challenging organisational problems, it will be worth paying $250 for an extra 1% certainty (assuming the problem is high-value / high-risk enough).

I’m not saying autopilot was a bad business decision - autopilot has fuelled sales and no doubt has been a great success and is amazing tech for being in consumers cars at fairly low cost.

I’m saying if the goal is to invent true robotaxis (because that’s the market you are going after) then I think waymo’s approach of following what works and cost-optimising later gets you there faster if you have enough funding.

Tesla Robotaxi is a classic case of premature optimisation - with Waymo they decided to deploy whatever technology worked best and did not cost-optimise at the start. Each Waymo vehicle is much more expensive than a Tesla, and has much more expensive sensors and compute.

Tesla on the other hand started with a load of constraints - smaller amounts of compute, cheaper sensors, needs to hit a price point as it will be installed in every vehicle sold - can't be $50,000 of dedicated self driving hardware per vehicle.

There are some scenarios where constraints lead to breakthroughs, but often for true-moonshot projects it is the opposite (see: Space shuttle, System/360, the manhattan project, LHC etc)

Not a bug necessarily, but one of the key features of trello is multi-user editing which doesn't seem to work.

i.e. if I open two tabs, and edit one tab, the other tab doesn't update and visa-versa until refresh.

Maybe - but it will still probably be less than a junior dev.

You could probably hammer the most expensive cursor API all-day every-day and it would still be a fraction of the cost of a junior dev.

oop found a security issue - you can abuse profile pictures to upload any file (e.g. to host malware if you were a bad actor!)

IMO don't think this is ready for production use in the slightest - but cool project!

A few other trivial bugs I have found:

* Can create multiple workspaces with same name which then ticks both * Invite user seems to not work randomly or will not send the email * Cards with special characters like @ will just not be created, and won't show error messages.

Just as a heads up, that roadmap you linked is broken.

The formatting looks all off for me on chrome mac (black bars) and then if I click on a card it opens a window but then doesn't load any data.

Also it looks like a bug that if you filter by some tags, then click a card, the filter gets reset.

This is 5 secs of testing on the one board you have publicly shared, so there might be a few bugs to iron out!

More cynically, they probably also know that the AI path is one that they can't win.

> Given the cited stats here and elsewhere as well as in everyday experience, does anyone else feel that this model isn’t significantly different, at least to justify the full version increment?

My experience is the opposite - I'm using it in Cursor and IMO it's performing better than Gemini 2.5 Pro at being able to write code which will run first time (which it wasn't before) and seems to be able to complete much larger tasks. It is even running test cases itself without being prompted, which is novel!

I think this seems to be veering off into some specific ethical viewpoint about milk supply chains and production rather than an analogy for product vs process.

But personally I can't see how the popularity of oat and almond milk in indepenent coffee shops tells us that much about how people perceive the inner workings of chatGPT.

I think this is a strange take - Technological advancements aren't mutual funds.

Of course it's an indicator of future performance - Not a guarantee, but certainly a indictator.

I like your analogy, but I don't think people even go "the cow gave this milk" - I think they tend to just go "mmmm yummy yummy milk"

People usually see the product rather than the process.

Bill Gates saw computers and extrapolated that there would be one in every home, Gordon Moore produced Moores Law which literally extrapolated compute in the 60s and almost still holding, and you can extrapolate that GPT-5 will perform better than GPT-4, and that there will probably be a GPT-6 that will be even better than that.

Extrapolating that technology will get better in the future when it has got better in the past isn’t a sure bet, but it’s a reasonably reliable one.

I don’t get how someone can look at these videos and think “wow there’s lots of flaws and it’s sloppy and no storytelling” rather than “holy smokes this stuff is improving fast!”

In 2 years we have moved from AI video being mostly a pipe dream to some incredible clips! It’s not what this is like now, but what will it be like in 10 years!

Totally agree.

Also the term 'LLM' is more about the mechanics of the thing than what the user gets. LLM is the technology, but some sort of automated artificial intelligence is what people are generally buying.

As an example, when people use ChatGPT and get an image back, most don't think "oh, so the LLM called out to a diffusion API?" - they just think "oh chat GPT can give me an image if I give it a prompt".

Although again, the term is entirely abused to the extent that washing machines can contain 'AI'. Although just because a term is abused doesn't necessarily mean it's not useful - everything had "Cloud" in it 10 years ago but that term was still useful enough to stick around.

Perhaps there is an issue that AI can mean lots of things, but I don't know yet of another term that encapsulates the last 5 years advancements in automated intelligence, and what that technology is likely to be moving forwards, which people will readily recognise. Perhaps we need a new word, but AI has stuck and there isn't a good alternative yet, so is probably here to stay for a bit!

I think AI is a useful term which usually means a neural network architecture but without specifying the exact architecture.

I think Machine Learning doesn't mean this as a word, as it can also refer to linear regression, non-linear optimisation, decision trees, bayesian networks etc.

That's not saying that AI isn't abused as a term - but I do think a more general term to describe the latest 5 years advancements in neural networks to solve problems is useful. Particularly as it's not obvious which model architectures would apply to which fields without more work (or even if novel architectures will be required for frontier science applications).

Sure, but everyone knows humans end up bringing down the database too by writing an innocent looking test query nobody else blinks at, which is why you end up needing a testing strategy for ANY SQL before YOLO'ing into prod.

If that’s the scenario, I would be asking why the testing pipeline didn’t catch this rather than why was the AI SQL wrong.

I think there is a world where you have both - a portion of research sitting under constraints, and also massive computing capability to help build the next big model.

Put another way - I doubt OpenAI would have built GPT-3 faster if they had less compute.

It’s not an either/or - it can be both!

If you have the best foundational research, but then find yourself blocked by compute, that’s a bad place to be in.

At the moment compute is a constraint too (at least for the smaller shops, maybe not OpenAI!)

We don’t necessarily know that the next AI breakthrough will need less compute either! A breakthrough that gives 5x performance for 20x training compute and 2x inference cost would still be a giant breakthrough and need a lot of hardware! Although this sort of breakthrough is unlikely, and it’s likely to be smaller steps forward - but I think it might be the case that models continue to get more computationally expensive even if there are breakthroughs that make models more efficient because we will want higher capability.

Why can't it? I think you may have slightly misread my post in terms of the single cell aspect.

> First to file would not allow Ocado to have patented the concept of the robot sitting above a cell/tote

Ocado patented the concept of a robot sitting above a single cell/tote, not the concept of the robot sitting above a cell/tote.

AutoStore's initial system was the first to have the robot sitting above a stack of totes, however the robot sat above two cells in it's original design (a 'cantilever design'). The settlement actually means that AutoStore cannot develop a robot sitting above a single cell/tote (note that their blackline robots sit across a 'tote and a bit' because of this - see image https://www.prnewswire.co.uk/news-releases/autostore-sues-oc...).

There is lots of historical complexity as with all patent cases, but the AutoStore loss in court doesn't mean that Ocado didn't copy them - it just means that Ocado have been judged to be legally allowed to copy them IMO :)

> Ocado would be the poster child [for being] an unscrupulous licensor "stealing" an invention from the actual inventor.

They absolutely are the poster child for this in the material handling world! This is widely known in the logistics industry.

I think the story is probably something like:

* AutoStore sold a solution to Ocado

* Ocado started modifying their own installation. AutoStore gave them the rights to do this naively thinking it was for their own use.

* Ocado started aggressively patenting their own modifications.

* Ocado then decides it’s going to build everything themselves, and also start selling their own version of the solution and compete directly with Autostore.

* Then the legal battle begins!

Broadly the main thing in contention is that Ocado patented the concept of the robot sitting above a single cell (ie tote). Autostore thinks this is obvious and shouldn’t have been allowed - their main reason for usually implementing a lower robot that sits across two cells is that it’s more reliable (lower centre of mass, simplified mechanics) but that they have now been blocked from just doing a smaller design of their original invention.

Autostore were probably naive at the time (they were still reasonably new to the automation market) and Ocado definitely had better patent lawyers - or at least as they were UK based had a better grip of UK patent law.

Ocado's first automated warehouse installation was also purchased from Autostore (before they built their own solution, and now market/sell it)

AutoStore losing a patent dispute doesn't mean that Ocado didn't copy them. Just looking at the patent dispute ignores that the first automated Ocado distribution centre was actually purchased from AutoStore, who had been selling their robot-digging tote system since 1996.

Ocado's initial patents as well were actually modifications of Autostore's robots, running on an Autostore grid, and Autostore manufactured the robots to Ocado's specification before Ocado decided to build the whole thing themselves.

So hard to argue that it wasn't a copy.

IMO I think the UK patent victory was a bit of a joke... Ocado's innovation of the robot above a single cell is both obvious, but also has it's own obvious downsides.

Not sure why Ocado gets so much credit for the latter though, they just copied AutoStore which has a fascinating history!

They purchased an AutoStore, then reverse engineered it, made a few changes, and claimed it as their own invention.

It depends what you count as learning - you told it something, and it then applied that new knowledge, and if you come back to that conversation in 10 years, it will still have that new knowledge and be able to use it.

Then when OpenAI does another training run it can also internalise that knowledge into the weights.

This is much like humans - we have short term memory (where it doesn't get into the internal model) and then things get baked into long term memory during sleep. AI's have context-level memory, and then that learning gets baked into the model during additional training.

Although whether or not it changed the weights IMO is not a prerequisite for whether something can learn something or not. I think we should be able to evaluate if something can learn by looking at it as a black-box, and we could make a black-box which would meet this definition if you spoke to a LLM and limited it to it's max context length each day, and then ran an overnight training run to incorporate learned knowledge into weights.

I'm not sure humans meet the definition here.

If you took the average human from birth and gave them only 'the most primitive first principles', the chance that they would have novel insights into medicine is doubtful.

I also disagree with your following statement:

> Right now we're trying to essentially train on the entire corpus of human writing. That is a defacto acknowledgement that the absolute endgame for current tech is simple mimicry

At worst it's complex mimicry! But I would also say that mimicry is part of intelligence in general and part of how humans discover. It's also easy to see that AI can learn things - you can teach an AI a novel language by feeding in a fairly small amount of words and grammar of example text into context.

I also disagree with this statement:

> One fundamental difference is that AGI would not need some absurdly massive data dump to become intelligent

I don't think how something became intelligent should affect whether it is intelligent or not. These are two different questions.

Why not tax the negative factors then, rather than the country of origin?

i.e. If the price is supposed to be a lever for labor conditions, why just tax China heavily and not Bangladesh?

Why tax more fuel-efficient European cars instead of American-Built Jeep Grand Cherokees?

And if reducing plastic waste is the priority, why would Trump's day include unbanning plastic straws?

Answer: It's not actually about reducing negative externalities, it's about geopolitics, otherwise it wouldn't be so negatively weighted towards a single actor.

Well AGI can write unit tests, write application code then run the tests and iterate - agents in cursor are doing this already.

Just not for more complex applications.

Code review does often find bugs in code…

Put another way, I’m not a strong dev but good LLMs can write lots of code with less bugs than me!

I also think it’s quite a “programmer mentality” that most of the tests in this forum about if something is/isn’t AGI ultimately boils down to if it can write bug-free code, rather than if it can negotiate or sympathise or be humerous or write an engaging screen play… I’m not saying AGI is good at those things yet, but it’s interesting that we talk about the test of AGI being transpiling code rather than understanding philosophy.