SgtBastard

Authentication is (proving) who you are.

Authorization is the process of determining if someone can do something.

That you’re conflating multiple distinct concerns doesn’t build confidence.

Temperature = 0 would always return the highest probability token at every generation step and given a single distribution of tokens at each generation steps,that would become deterministic.

Where that gets murky is… perturbation inside the models themselves, such as Mixture of Experts (MoE) models who’s more internal parameter activation is not so deterministic (and therefore the tokens they generate are not deterministic across runs)

Are you asking “is the probabilistic inference model 100% deterministic?”

If so: no.

Just leak them to the inference providers, obviously /s

If you have self hosted models and/or self hosted APIs, maybe you don’t need MCP to provide a gateway to a secure resource.

If neither of those things are true, you need an authenticating gateway/proxy or a target API that supports single use credentials (and get the model to generate a call to use them).

We can argue whether MCP is a good authenticating middle layer, but not whether one is required.

I always have a Mandela-effect moment about the Dropbox and iPod dismissive comments with Hackernews and Slashdot respectively.

God I’ old

I asked the original commenter to confirm my understanding of what they were saying.

I find it a fascinating alternative view to what is largely well understood (your counter-point).

The thing that stood out is the comment that it’s an approximation of the original data generator (humanity). Early approximations were poor (GPT 2-3, to an extent GPT-4).

I’m not so sure I can reject the hypothesis that such an approximation can be found.

> A lot of people seem to think of the training process as curve-fitting data (the "stochastic parrot" model), but I think of it more as "solving an inverse problem to approximate the unknown source that generated the training data"

I just wanted to confirm your underlying point here: training a model isn’t about finding a function that fits the observed data (even though that’s the outcome) but instead finding an approximation of the unknown source that generated the source data in the first place.

In the case of LLMs an approximation not of a human but of the sum of humanity that produced the training copora?

You may be discounting the tokens generated in the thinking trace but not included in the output to users.

The “problem” is exactly as you’ve framed it: individual 1-1 tutoring unambiguously improves education outcomes by a significant (2 StDev’s in Blooms study) vs any other understood method but delivering this universally is (was?) infeasible.

Bloom looked at other methods used in concert to achieve similar improvements to “solve” the problem that could be delivered at scale.

LLMs may provide a new path to the 2-Sigma improvements without the same delivery problems.

Thank you for this and for parent’s comment - I know what rabbit hole I’m going down today.

Mistral in the EU, for one.

Mistral (a French company) shouldn’t be discounted.

What makes you think 1-2, 6-8 can’t be done by agents?

Yes - itself.

Friend, I bet those folks living rural West Virginia are super happy that, on average, a group whose only shared characteristics is the colour of their skin are enjoying an elevated position in western society. Super happy. All racism is gross.

How do you feel about your current levels of dakka? </40k>

Are you ok there? SAML, OIDC and a depressingly long tail of Kerberos is how modern enterprise identity security works.

We have multiple threads of research demonstrating in-context learning, friend.

https://github.com/dqxiu/ICL_PaperList

This thread is about coding agents, of which a model is only one (important) part.

Fancy while() loops is how I describe them.

Claude, with a modicum of guidance from an engineer familiar with your monolith, could could write comprehensive unit tests of your existing system, then refactor it into coherent composable parts, in a day.

Not doing so while senior management demands the use of AI augmentation seems odd.

Only if you’re relying upon the models to recall facts from its training set - intuitively, at sufficient complexity, models ability to reason is what is critical and can have its answers kept up to date with RAG.

Unless you mean out of date == no longer SOTA reasoning models?

Guilds of Knowledge hoarders and AI adepts are my bet as well.

Thank you, I learnt a new term today.

… the internet was not immediately useful in a million different ways for almost every person.

Even if you skip ARPAnet, you’re forgetting the Gopher days and even if you jump straight to WWW+email==the internet, you’re forgetting the mosaic days.

The applications that became useful to the masses emerged a decade+ after the public internet and even then, it took 2+ decades to reach anything approaching saturation.

Your dismissal is not likely to age well, for similar reasons.

Nitpick: Language is technology, it’s not something we’re genetically born with and is critical for education to happen.

Technically you are correct but the commenter you’re responding to means that with the amount of Western Governments spend on MS products and services, the are a d facto (if not de jure) state backed enterprise.

Only if you profoundly misunderstand what GC is.

Ping Identity run the largest enterprise identity platforms on the planet after merging with ForgeRock last year. Think HSBC, JP Morgan Chase-scale.

>Humas learn dynamically

In-context learning (Few Shot Prompting) is absolutely demonstrable by LLMs since GPT 3.5 at least.