i.e reinforcement learning against a weak reward function - benchmark is insufficiently complex and is not representative of the real world sufficiently.

The "game", i.e. decision tree can be modeled as a multi-arm bandit problem, to deploy finite resources ( compute) toward exploitation/exploration .

The main issue is each training / fine-tune is very expensive so number of chances at the slot so to speak is pretty limited today.

It would be feasible to change something like that without breaking security now.

Google can hardly start allowing/routing a new account for first.last@gmail.com when you were getting it for years even though your account is firstlast@gmail.com and sensitive communication like say from your bank would routed there.

In my experience most SaaS apps do not filter this out and allow re-sign ups with sub-addresses.

Gmail has an additional behavior that dot character is ignored in local component of the address . multiple@gmail.com, mult.iple@gmail.com mult.ip.le@gmail.com all route to the same inbox as well.

[1] https://datatracker.ietf.org/doc/html/rfc5233 [2] Less common in work hosted ESPs but almost universally default enabled in public ESPs for consumers.

The main llm will refuse to scan for issues flagged or not, and the cheap model not do a good enough scan on its own.

For models designed/marketed for cybersecurity defensive uses, any predictable refusal mechanism is a vulnerability. It is like being able to cause a kernel panic or segmentation fault .

Even if the gate is fail-reject, an attacker can overwhelm HITL reviews with many false positives and use DoS vectors here.

> a human would generally spend a lot of time here,

This was always the hard part - game engines, asset libraries and all other services / SDKs were always making code-generation cheaper every generation.

Releasing a bug-free, thoughtfully built product requires a lot of attention and product skill.

Same reason developers continue to use Apple payments even when they have to shell out 30% of the revenue.

I can see Apple, setting App store rules around declaring AI usage, or could start labeling apps not using their models with strong language designed to amplify the increasing user concerns around AI and so on.

The product strategy has to be better the product itself does not have to be objectively better for the developer for them to have to choose it.

Backrooms was a quite successful web series on YT which in turn originated in 4chan boards.

Only the medium being sourced from is changing from successful Broadway shows, popular novels or comic books in the years past. The calculus remains the same - properties with name recognition even from other formats tend to be green-light.

The concern basically boils down to how large and serious is the team and what if they abandon the project in few weeks or months .

These were always the risks, many here have been burned by betting years of their career building against promising but what turned out to be weak projects

OP is alluding to the fact that today commit frequency, lines of code or how active the contributors in the issue trackers are no longer good signals to use as proxy.

When the underlying project to yours is few million lines of code written by machines only it is not going to be feasible fork and maintain or in-house it if the maintainers abandon it

To be clear users of a library or a tool aren’t owed anything when it available gratis and fully open source .

However not everyone has access to unlimited tokens to disregard the quality (in terms of history and usage ) or size of the underlying project completely

Any views they have on this topic is going to come across as quite opinionated given their choices for text rendering for this post and general aesthetics of website.

Using the keyword “Workflow”like “Ultrathink” is problematic?

Ultrathink is uncommon enough that it is unlikely to be used in code or prompt outside its intended purpose.

Workflow is generic keyword and used in so many contexts both inside the codebase and orchestration tooling like say temporal.io or others that name their constructs “workflows”.

Everything else is subjective to your setup, use case, configuration tuning and so forth.

More importantly bean-counters and decision makers at even 150+ seat orgs are looking at pricing sheets and enterprise contracts not how it performs for some team in a specific harness today to make million dollar annual contracts. It is not common for procurement teams to do commission the level of detailed analysis or large scale pilots that will actually hold for the duration of contract.

That doesn't mean that GPT-5.5 is selling less than Claude at all, just that cost is not the primary driver if list price is not cheaper, there is reason these are published in the same format by every vendor, because the common metric is how finance likes to compare with.

Input/Cache/Output ratios are use case and configuration dependent . Any benefits in one model can usually be roughly to another with configuration tuning, and discussions devolve into subjective experience.

Pricing sheet is the objective way to compare cost.

Everyone has critical risk on multiple parts of the supply chain. GPUs and Memory are just things OAI mitigated for.

Power - Bigger bottleneck than GPU or RAM perhaps, New Grid connected capacity is typically 10+ year timescale with lot of regulatory friction. Captive capacity is also quite constrained - now Gas turbines have 7+ year wait time.

There are plenty of hard constraints that OAI cannot easily solve either.

GPT-5.5 is a bit more expensive than Opus ? Current list prices

  | Model      | Input   | Output   |
  | GPT-5.5    | $5/MTok | $30/MTok |
  | Opus 4.8/7 | $5/MTok | $25/MTok |

> 10.4 years of production coverage

Kinda true, airlines and manufacturers like to do big order announcements/deals for their future needs of few years all upfront. If Airbus suddenly delivered all 9k aircraft most airlines simply cannot afford it, or take possession and use them even.

For example Indigo is Airbus only operator with a fleet of 450 today and has around 920 more Airbus aircraft (10% of the book) on order. Neither Indigo or Indian aviation sector( of which Indigo is 60%) can triple the capacity today . India need serious upgrades (Terminals, Runways, Gates, new airports) coming online and also demand maturing, i.e. more people can afford to fly for that kind of volume to make sense which even the best scenario will happen over the next decade.

For more mature/slow growing airlines it is function of existing fleet age and the optimal point each aircraft is retired/sold , doing it too early will make them unprofitable .

It is a less a backlog and more their next 10 years of committed sales.

P.S. There is whole other industry aspect around Buy-Sell-and-leaseback financial engineering that can drive order volumes a bit. The backlog/order book also have commodity futures aspects.

Every generation of developer tooling that increase of absolute code throughput creates a new class of developers (and users).

Always been the case since first compilers, through eras of frameworks to today, and the skill level needed to be one has dropped. In mid/late 80s only Master / Doctorate level Comp Sci professional could write any applications. It dropped to undergrad and just Information Technology engineers and comp sci theory became mostly optional and dropped further to any college level educated with some training and has been trending below with no/low code tools like retool pre 2022, that was before agent codegen services such as v0/replit and so on.

The next generation developers will not produce applications and architecture as previous generations did, just as we most of us here don't produce the level of quality that pg did when building this platform[1] , but as long as the user can find value it doesn't matter as countless enterprise applications of middling quality already prove today.

All this to say the 200M/30M numbers will not remain the same is the thesis for these businesses, will it change by large enough at a fast enough pace to justify the capex, I don't think so either. However web 1 then 2.0 , saas and mobile revolutions were pretty quick with new class of users and developers so not completely unrealistic .

[1] While HN is a heavy outlier with its custom lang lisp implementation, there are any number of examples from previous eras that are more moderate in choices but written with solid architecture with skill levels would be hard to find in today's generation founders.

> Play money means it's much easier for anyone anywhere in the world to get started and try out forecasting without any risk. It also means there's more freedom to create and bet on any type of question.

Their investors/lawyers probably did not want to back the risk using real currency adds.

> they wasting - and why?

i18n language models are not area something frontier labs are focusing ton of resources on? ( certainly not in Norwegian)

The corpus of content in Norwegian - may not require very large clusters, or even if it does, this is best that the library could do, it would be certainly more than anyone else is investing in Norwegian models

SOTA models do not have the access to the quality of content that the national library does? The article mentions licensing with newspapers specifically, and the library has access to its own content archive.

English and Norwegian are not closely related language families, perhaps LoRA is not best approach?

I am curious if there is published research on how well localization works with LoRA depending on how far off the target language grammar/vocabulary is from English.

Projects like this typically have more than one objective and are not only building SOTA project, but is also to build/train foundational local talent , similar to universities launching satellites .