rhdunn

It depends on whether you are using the LLM as a) an advanced autocomplete; b) a chatbot/query interface; or c) an agent. Only the latter goes and does things for you.

It's also confusing that the term AI has become synonymous with LLMs as it covers other things like text prediction/correction.

So AI can be a tool or a service depending on how you are using it and what you are using it for.

These models aren't really LLMs -- they don't just operate on text tokens. They often include vision models and in some cases audio models. That means that they can better associate the meaning of images and words together so that when someone says "make this button blue" or "create a 3D model of a rocket" they have some level of understanding of what that is and what needs to be done.

The key question is how good that understanding is. For example, a model would likely have a good understanding of various named colours and hex values (e.g. from the HTML specs, X11 specs, and various colour comparison websites) such that it could reasonably correlate that to a CSS entry. It's not clear if/how well a model would identify that given an image, though it should be easy to generate a dataset of image to colour name and/or hex code for training and evaluation.

What's more interesting is whether these frontier models are at their core transformer models, whether they use residual streams to facilitate learning, and whether they are using some other as yet unpublished architecture that gives them an edge.

You're comparing a company (Shopify) to an application (Google Chrome). A better comparison would be the number of engineers at Google (90,000 [1]) vs Shopify, or the number of developers actually working on the mobile application instead of any other area (website, backend database, internal systems and processes, infrastructure management and maintenance, etc.).

[1] https://www.linkedin.com/posts/rpandey1234_its-mind-bending-...

I'm not an AI critic (I use it for autocompletions and asking some questions) but still write my own code.

I'm not sold on the agentic workflows that AI labs are pushing and haven't yet figured out if/when/how to integrate it into my workflow.

There are two general things I'm weary of with agentic workflows compared to the chat-based workflows:

1. the agentic workflows are much more inclined to go and do the thing rather than involve you in the loop -- e.g. if you are trying to design/plan something;

2. the agents are happy to go and run any command -- you can get them to prompt to confirm the action, but they could easily do something like wipe your home directory, install a random package, or something else.

So this would suggest that the Goidelic languages (Irish Gaelic, Scottish Gaelic, Manx) are derived from the Brittonic languages (Welsh, Breton, Cornish). [1]

I wonder how this affects the different proto-Celtic to modern Celtic languages hypothesis. [1]

This would suggest the following language tree: Breton -> (Cornish / Welsh) -> Irish Gaelic -> (Manx / Scottish Gaelic).

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

Ruby tags apply to other scripts as well like Japanese and Korean. See https://www.w3.org/TR/css-ruby-1/.

It only makes sense to train a frontier model if you are trying a different architecture to one that is available from an existing frontier model. This is because the different model architecture will learn the weights differently.

It may make sense to train a frontier model on an existing architecture if the base model is not available and the instruction trained version doesn't fit with what you want. There are techniques like ablation, but those could have other effects on the model, and there can still be lingering effects of the instruction training in the model that surface less frequently (e.g. on an input not covered by the ablation training).

Otherwise, fine tuning is definitely the way to go. However, you need to be careful not to over-tune the model such that it is only tuned to the data you are training it on.

My understanding is that 1) SR doesn't consider acceleration (it's an extension of Galilean/uniform motion), and that 2) when Einstein considered acceleration as well as gravity via the equivalence principle which lead to GR [1]. The key insight of the equivalence principle was that the force from gravity (e.g. standing on the Earth) is no different to the observer in their frame of reference to them being in a room in a rocket accelerating at the same rate as gravity [1], [2].

Thus, if you extend QED/QCD/SM in a similar way (thinking of QED/QCD/SM extensions in terms of acceleration and curved space with the equivalence principle in mind) that may lead to a quantized theory of gravity. -- Sir Roger Penrose has a similar idea/thinking [3].

One of the key challenges with quantizing gravity is in how the terms in the expressions resulting from analyzing the Feynman diagram interactions behave [4] which prevent them being renormalized. For electromagnetism you can formulate the terms using the fine structure constant (via the coulomb potential, ħ, and c) which results in successive terms decreasing in value and thus stabilizing to a single value.

For gravity using Newton's relationship between two masses in a similar way to deriving the fine structure constant you get Gm^2/ħc. Applying E=mc^2 gives GE^2/ħc^5. Using the Planck energy constant gives (E/E_p)^2 for the energy coupling strength. This means that unlike electromagnetism, the successive terms in the Feynman diagram analysis grows exponentially instead of decreasing to 0. Thus, this approach to quantization doesn't work for gravity.

Note: you can still use this to analyze quantum gravitational effects at small energies by evaluating to a given number of terms.

[1] https://www.britannica.com/story/how-albert-einstein-develop...

[2] https://www.ebsco.com/research-starters/physics/equivalence-...

[3] https://www.youtube.com/watch?v=VQM0OtxvZ-Y "We need to 'gravitise' quantum mechanics, not quantise gravity | Roger Penrose | Full interview"

[4] https://www.youtube.com/watch?v=yTEPm5d6mrI "Why Quantum Gravity Doesn't Work"

The article says that this proves that Einstein's equivalence principle (resulting in relativity) holds in this test of a falling quantum particle (where gravity results in a phase shift in the quantum state).

It doesn't show/prove how general relativity and quantum mechanics interact.

NOTE: The Dirac equation and Quantum Electro Dynamics (QED) unify quantum mechanics and special relativity (non-accelerating frames of reference).

So the remaining piece is either to extend QED/QCD to accelerating frames of reference or to quantize general relativity. That would likely predict the phase shift observed in this experiment.

Welch Labs on YouTube has a great collection of videos on how AI models learn. His recent video [1] covers how image models can learn to encode reasoning in the image processing layers when not given an out of band reasoning set of weights to use instead. I suspect that this applies to LLMs and CoT reasoning vs output token weights.

[1] https://www.youtube.com/watch?v=QgH9sr7G13Q

1. Does the implementation fit in the architecture/style of the project?

2. Are there potential security, accessibility, performance, etc. issues?

3. Domain specific knowledge (SQL, ASP.NET, XQuery, etc.) where there are better ways of solving a problem, or possible issues not handled.

4. Sense checking ... is the code easy to read? does it need an explanatory comment? does it need named parameters? etc.

My theory behind this is that when you encounter something new/novel it is distinct enough to remember -- where you were when you learned of 9/11; the days around the start of covid/lockdowns; being in a car accident; getting married; etc..

However, when you experience the same thing repetitively -- having breakfast; getting ready for school/work/etc. -- it doesn't make sense for your brain to remember each of these events as distinct events but to mush them together into a single or combined memory. This may be due to how the brain indexes/references memories and that memories with the same references get combined together.

This is also partly why time appears to go faster when you get older -- the same day-to-day events blend together but the key milestones (holidays, etc.) stick out.

It is clever, but makes comparing versions of the WordNet database a lot harder as the byte offsets change as entries are added and removed.

"Secret Hitler" is a social deduction game (https://en.wikipedia.org/wiki/Secret_Hitler). You can find various games of it on YouTube.

The reasoning is that 600 is 60x10 which is conflating base 60 and base 10 numbering systems. You would expect something like 3600 (60x60), or even (given things like our units of time) 720 (60x12). The latter is because 12 is a factor of 60 and is easier to work with in that number system/base.

Keyboard shortcuts for those menu items are what most people will learn -- save, print, cut, copy, and paste will be the common ones that a lot of people will know.

I'd be willing to bet that professional typists and people who spend the majority of their day typing (email, word, IDEs/text editors, etc.) are the ones who will know more keyboard shortcuts and navigation, especially for the things they do repeatedly.

In some cases you may need the keyboard accessibility.

- Your mouse/trackpad breaks or isn't recognized?

- The right mouse button for your trackpad isn't working for some reason (e.g. you spilled a drink on it). It should still be possible to use the context menu keyboard button to bring up the context menu.

- Your hands are on the keyboard (typing a document, email, code, etc.) and it slows you down to move to the mouse/trackpad to perform an operation such as copying/pasting text or bringing up the spell checker context menu actions so you want to do those with the keyboard to keep yourself in the flow.

Most often you'll learn keyboard shortcuts/interactions for your workflow then expand out as and when you need them. Sure, knowing all shortcuts and interactions is fairly limited, but the majority of people will know how to cut/copy/paste with the keyboard as those are common operations.

My go to when implementing custom controls that can't be achieved in native HTML is the ARIA Patterns (https://www.w3.org/WAI/ARIA/apg/patterns/) site. That has clear documentation on the roles, states, and keyboard interactions for common UI controls. I've used it for things like a search bar with an autocomplete dropdown to get the keyboard interaction right.

That's why you don't just use colour for identification. E.g. include a warning icon next to a button/element in addition to using a warning colour.

It would be more flexible for a compiler to reuse the range analysis logic used in optimizations for statically verifiable divide by zeros. That way you could extend it to other things like statically verifiable overflows.

Will they send him ... TO SPAAACE?!

ASM is still used in some areas, such as:

1. function trampolines, such as in libraries like ATL (Microsoft's COM template library);

2. hand-optimized encoders/decoders for things like ffmpeg;

3. niche applications such as Ben Eater's 6502 breadboard computer (using a modified WozMon and Microsoft BASIC);

4. reading ASM output from C/C++ code in things like the GodBolt site for performance comparisons/analysis;

5. reverse engineering old DOS games;

6. people working on things like compilers (generating optimized assembly), JIT engines, or other code generation -- though this will typically be machine code, possibly with assembly instruction comments;

and other performance critical places where assembly can speed up that part of the code.

You mean like Microsoft, Google (GMail, YouTube, Android), Apple, Facebook, and others?

Sorry, you can't buy an iPhone because Apple has got too many customers.

Sorry, you can't have a GMail account because Google has too many customers.

I had my GMail account locked for 1-2 years because I accessed it from my parents house (in the same country but in a different county) while on holiday. That was because they detected the account being used from a different IP address.

Using VPNs can also trip this.

Part of it is scale. If you have a small number of clients/users it is possible to provide that support. If you have millions or billions of users then you can't scale the support to handle the support requests, so some form of automation becomes inevitable.

The cost of training a model from scratch is going to be cost prohibitive for the vast majority of companies (even if renting the hardware needed for the 1-2 month training time). It's an interesting learning exercise, and some of the things learned can be applied to other parts of the process. There's also the issue of needing a huge amount of data needed to get decent weights.

Fine-tuning a model or LoRA based on the companies data set is more feasible but you're likely going to need several runs as you test/try out different base models, parameters, etc. This is why there are a lot of fine-tuned models on huggingface based on base or instruction-trained models from the larger AI companies that have released open weight models (Microsoft, Google, IBM, Mistral, DeepSeek, Qwen, etc.).

Training is limited on memory first (storing training data and weights) and computation second. Realistically you need to own or rent 2-8 H100/B100 devices or Google's TPUs.

The majority of workflows for a company providing AI capabilities are likely best solved by tailoring a system prompt for the chosen model, evaluating the prompt and model with tools like promptfoo, and then running it on a compute cloud provider (including AWS Bedrock). If the company is big/financially well off enough they could look at buying the hardware needed to run it on their own servers.

For other uses like agentic software development you'd need to spin up a suitable model on a compute cloud provider (or local hardware if the model is small enough) and then tell your IDE/editor to use that model. You would need some way of benchmarking and evaluating the models to see if they are capable of doing the tasks you need. -- There have been some tests done by people on YouTube that suggests that Qwen 3.8 27B is a decent model, but your needs may vary.

20 years ago is 2006. I'm pretty sure I had a digital camera (DSLR) around that time with enough resolution (10 megapixels, IIRC) for decent photos. Wikipedia notes that the Canon EOS 1D Mark III [1] was released in 2007 with 10MP.

According to wikipedia [2] the first consumer DSLR camera was in 2000.

[1] https://en.wikipedia.org/wiki/Canon_EOS-1D_Mark_III

[2] https://en.wikipedia.org/wiki/Digital_single-lens_reflex_cam...

Digital cameras have existed before phones had the capability and have been around for a long time (at least 20 years). They made it a lot easier to take photos as you can take hundreds/thousands of photos and download them when you get home.

When visiting places on holiday, I've taken my camera with me for the purpose of taking photos. I've also seen people with more expensive cameras (mainly larger lenses) camped out near puffins, on boat trips, etc. with the intention of getting a decent wildlife photo.

95 years after publication. Many other countries also have an X years after the author's death clause where X varies between countries but is at least 70.

There are also other weird issues such as the UK having a clause protecting Peter Pan (so a children's hospital gets royalties) and the King James translation of the bible (under Crown copyright) that extend the copyright even further.

In short, it's a mess.

Audio transcription is a hard problem that is inherently non-deterministic and probabilistic due to ambiguities in the speech due to:

1. accents -- Especially around mergers (cot-caught [AmE], trap-bath [BrE] vs palm-bath [LondonE], pin-pen [Some AmE]). These can even be hard for native speakers -- try transcribing a broad Scottish, London, Brooklyn, or Indian accent and see how well you do.

2. sound/phoneme variation based on surrounding phonemes -- It is common for the 'n' sound to be realised as an 'ng' sound before a 'k' or 'g' sound due to velarization ('ng' is the velar variant of 'n' and 'k' and 'g' are velar sounds). It is common for vowels to be nasalized before nasal sounds ('n', 'm', 'ng'). It is also common in non-rhotic (don't pronounce the 'r's next to vowels like in 'start' and 'north') to pronounce an 'r' between two adjacent vowels in words ending/beginning with vowels (the "intrusive r", e.g. in "there and back").

3. sound changes due to fast speech ("I'm gonna see 'bout it t'day.", etc.)

4. ambiguity about where words start/end (e.g. "to Damon" vs "today mon" where the "mon" is the variant of "man" in Caribbean English).

5. word play, puns, etc. due to accent and other speech.

6. technical words in a given domain, specific place names, etc.

7. other things that can affect speech such as mumbling, stuttering, or slurred speech.