bravura

For example, in difficult perceptual tasks ("can you taste which of these three biscuits is different" [one biscuit is made with slightly less sugar]), a correlation of 0.3 is commonplace and considered an appropriate amount of annotator agreement to make decisions.

I never said that I support this oversimplification of Portugal. Which is why I live in a small town, socialize with Portuguese, and study Portuguese.

I've tried prunes operations while preserving referenced components using redocly.

I've tried openapi-extract CLI to extract only components I reference.

And I've tried openapi-format CLI.

They all give different results, and I can't tell whether I am pruning too much or too little.

And yes I'm using spectral at the end, but it doesn't necessarily show it you're missing something that isn't referenced in the final output.

Any recommendations?

Source: I live in Portugal and notice all the discussion about it from my Bay Area friends who want to expatriate. To here, specifically.

LeCun's argument is this:

1) You can't learn an accurate world model just from text.

2) Multimodal learning (vision, language, etc) and interaction with the environment is crucial for true learning.

He and people like Hinton and Bengio have been saying for a while that there are tasks that mice can understand that an AI can't. And that even have mouse-level intelligence will be a breakthrough, but we cannot achieve that through language learning alone.

A simple example from "How Large Are Lions? Inducing Distributions over Quantitative Attributes" (https://arxiv.org/abs/1906.01327) is this: Learning the size of objects using pure text analysis requires significant gymnastics, while vision demonstrates physical size more easily. To determine the size of a lion you'll need to read thousands of sentences about lions, or you could look at two or three pictures.

LeCun isn't saying that LLMs aren't useful. He's just concerned with bigger problems, like AGI, which he believes cannot be solved purely through linguistic analysis.

The energy minimization architecture is more about joint multimodal learning.

(Energy minimization is a very old idea. LeCun has been on about it for a while and it's less controversial these days. Back when everyone tried to have a probabilistic interpretation of neural models, it was expensive to compute the normalization term / partition function. Energy minimization basically said: Set up a sensible loss and minimize it.)

A lot of applied work in vision and audio glues together different existing modules, instead of training the whole thing end-to-end.

In an ideal world, things are ideal. But the world isn't a grad student's wet dream.

But there are not as many influential papers in the world as there are people who respond: "Why didn't they train the whole thing end-to-end?"

Jobs was talking about how people don't know what they want until you show them. He was fighting in reaction to the idea that what people say they want will guide you in the right direction.

'A studio version was made for John Peel's BBC Radio 1 show at the Maida Vale Studios on 28 October 1977 and first broadcast soon after on 14 November. It includes a separate section that begins with an abridged session version of the original comprising the first verse, chorus and guitar outro - this part has become known as We Will Rock You (slow). Between the two parts there is a brief reading of Hermann Hesse's novel Siddhartha, used in a BBC Radio documentary. This audio was found on the BBC tapes being reused to record the session and was retained by the band.'

Germany: Requires extensive justification for firing, works council approval, and mandates reinstatement for wrongful termination. Employer bears heavy burden of proof.

France: Must prove "real and serious cause," follow strict procedural steps, and provide mandatory tenure-based severance. Special protections for worker representatives.

Netherlands: Unique system requiring pre-approval from government agency or court before any dismissal. Heavy emphasis on severance.

Italy: Article 18 mandates justified grounds (economic or disciplinary) for firing. Court-ordered reinstatement common, especially in large companies.

Belgium: Requires detailed dismissal grounds and long notice periods up to 15 months. Mandatory outplacement services.

[edit: Also, you can always move back to the US and get a new job.]

* You persuade your manager that this is okay and name the country.

* BigCo Inc. terminates you and BigCo Lda. or Srl. or whatever there subsidiary in that country makes you an offer letter.

* Your base salary in the new offer is 1/3 less, because "the cost of living" is less. If you ask them if you should do 1/3rd less work, they give you a puzzled look. If you ask them if they discount their prices to customers in low COL countries, they will get mad.

* The non-cash part of your compensation (stock, etc.) should not get impacted much, except that when you get taxed you'll pay more tax. Money isn't everything and you're obviously contemplating this move for other reasons.

* You will accept the offer and follow the immigration procedure of your new country. BigCo will have immigration lawyers they retain to ensure they can hire the candidates they desire, for example you.

* You will chill for a little time, maybe a month, while the immigration process goes through.

* Congratulations, you now have a work visa, benefits, and all the privileges that being a legal employee in that country confer upon you.

Slightly worse are arrangements in which you work through a subcontracting company, don't get benefits, and/or have to fight for the visa yourself.

The other ideas you are contemplating probably don't work, and if you don't understand why you should ask a chatbot not HN.

"Aider now has experimental support for using two models to complete each coding task:

An Architect model is asked to describe how to solve the coding problem.

An Editor model is given the Architect’s solution and asked to produce specific code editing instructions to apply those changes to existing source files.

Splitting up “code reasoning” and “code editing” in this manner has produced SOTA results on aider’s code editing benchmark. Using o1-preview as the Architect with either DeepSeek or o1-mini as the Editor produced the SOTA score of 85%. Using the Architect/Editor approach also significantly improved the benchmark scores of many models, compared to their previous “solo” baseline scores (striped bars)."

In particular, recent discord chat suggests that o3m is the most effective architect and Claude Sonnet is the most effective code editor.

Would that suggest to you what you did wrong? Or purely show you what you got right? How chatty is this variable?

This is the wrong way to ask for help.

Instead, consider offering your help and time apprenticing and learning along the way. Can't code that well? Write test cases and clean up. Or help blog writing. etc. You certainly have some valuable skill you could trade up.

For a niche product like this, most likely it will be hot pile of garbage.

e.g. you live in EU but the account home is the US, or you live in Portugal and the streaming provider thinks you're in Spain, etc

Abstract: The advent of Transformers marked a significant breakthrough in sequence modelling, providing a highly performant architecture capable of leveraging GPU parallelism. However, Transformers are computationally expensive at inference time, limiting their applications, particularly in low-resource settings (e.g., mobile and embedded devices). Addressing this, we (1) begin by showing that attention can be viewed as a special Recurrent Neural Network (RNN) with the ability to compute its many-to-one RNN output efficiently. We then (2) show that popular attention-based models such as Transformers can be viewed as RNN variants. However, unlike traditional RNNs (e.g., LSTMs), these models cannot be updated efficiently with new tokens, an important property in sequence modelling. Tackling this, we (3) introduce a new efficient method of computing attention’s many-tomany RNN output based on the parallel prefix scan algorithm. Building on the new attention formulation, we (4) introduce Aaren, an attention-based module that can not only (i) be trained in parallel (like Transformers) but also (ii) be updated efficiently with new tokens, requiring only constant memory for inferences (like traditional RNNs). Empirically, we show Aarens achieve comparable performance to Transformers on 38 datasets spread across four popular sequential problem settings: reinforcement learning, event forecasting, time series classification, and time series forecasting tasks while being more time and memory-efficient.

I also need an SDK to script it, tbh. What I want is to have actually a few different agents that interact with each other. Do you expose a good SDK?

Okay, so who's doing this today and how?

This question came up recently in the Aider discord and not many had a good answer.

Aider is great but the SDK is weak and second-class, so interacting with the repl frustrates agent-dev.

Sidecar (which can run independently of Aide IDE, not to be confused with Aider), https://github.com/codestoryai/sidecar/ is one agent that was mentioned. Many of that projects issues are auto-responded to by a PR-creating agent.

Anything else I'm missing?

In general, I know how I would build an agentic dev-loop, I'm just looking for a good SDK that handles prompting and diff merging etc. i.e. Aider as an SDK or similar.

Latent space reasoning can represent and manipulate UNCERTAINTY more concisely and elegantly than token space reasoning.

Am I the only person in this world who is an accomplished engineer and thinks commit messages are worthless? That if I want to roll back or dissect, I do it at the PR and not individual commit level?

The need for very accurate comment messages occurs so rarely in my workflow that the value of crafting them correctly is not there.

I ask honestly: Am I missing something here? Why? Is it something peculiar to my workflow I'm missing?