scottcodie

They're definitely not the only one. I've been building on relational transformers, which does prediction and classification over relational data (it handles numeric types better). It's validating to see that these small models that do prediction tasks are so useful to the community, but also stings a little that it was so hard for me to communicate how game changing they are.

You can also mine attention from image models, it's a lot of fun and very interesting.

A new transformers architecture based on relational data instead of text. Been exploring search reranking based on graph data and customer prediction.

https://github.com/relativedb/relational-transformers

https://relationaltransformers.com/

You can read more about the architecture here: https://relativedb.com/blog/relational-transformer

I did took a native c++ approach when writing a relational transformers engine (RelativeDB). My journey was pytorch -> c++ -> Triton (lang). While C++ was more performant than Triton, I couldn't afford to optimize on every gpu. I just accepted the ~15% throughput loss for my cloud service, which honestly wasn't bad for the amount of flexibility I got out of it.

But the cpp port of vllm looks great, that'd be great if you'll maintain that. I hit the same limitations with vllm.

The whole graph is walked at inference time to create the context. There is an attention feature, called 'neighbor attention', that propagate signals along primary-foreign key relationships. The task table includes the foreign keys themselves.

I created a prose tool as a claude code plugin to catch these and other AI-isms. It's not really intended for release quite yet, but it has been very useful. https://github.com/MariHQ/mari-cc

If this was a claude plugin with a hook on my dep files, I'd be in.

I'm not sure a game needs transactions. It seems to me that entity streaming concepts are a better fit: reading from an event queue and consistently reconciling the state over time. Like the sorted top-n in the example, if it was a game and the query is generally known, it would be better to just materialize it as a fast rank, which goes beyond the IVM-style materialization they suggested.

Have you taken a look at the various test harnesses for video encoding/decoding, or consider publishing a standalone one yourself? I get a bit-for-bit rewrite can have some advantage but it's hard to know if the AI implementation actually covers all of the edges cases.

I have a legit laminar flow hood (airclean 600, 32 inch) in my home for plant tissue culture. I gotta say, I thought it was going to be easy street to sterile culture- it was not. Sterile protocol matters so much that I think it would have been just as easy to do what I needed to do in a glove box with all the extra precautions I had to take. It's honestly hard to beat an enclosed container with zero airflow.

One thing the other blog post missed and this post misses too is that you don't need Kafka to use Debezium with Postgres. This gives you a pretty seamless onramp to event streaming tools as you scale.

I've spent my entire career developing databases (oracle, cassandra, my own database startup). Knowing if your workload is read or write heavy is one of the first questions when evaluating database choice, and is critical for tuning options. I would give this article hate just because it feels partially written by AI and the title needs a possessive 'your' in it, but its core ideas are sound and frame the issue correctly.

Stream processing/materialization engine written in rust that can be compiled to wasm.

Graffiti art.

Is there a definition of a "security update"? Software has an infinite number of bugs and it is cost infeasible to fix them all. If it's years down the road, the engineers that wrote the code may be long gone.

Paying poorly would bias you to candidates who will accept poor wages. Good candidates, even if they are young and inexperienced, can have a high market rate.

You could also use a temporal table.

SQL in the context of a single query struggles with a composable features. But looking slightly outside the scope of queries with assignment statements and view/materialization dags then you start getting some of that composition back. I think SQRL is an good example of SQL with a bunch of composable feature bolt-ons.

https://www.datasqrl.com/blog/sqrl-high-level-data-language-...

Most technology advances are not labor replacing but rather labor augmenting. For example, LLMs could make teachers much more productive in the classroom but it would be unlikely to replace teachers entirely.

Price gouging is often seen as immoral because it involves high prices without a justifiable reason. It is often considered in absence of welfare analysis, even if the middle class can afford the price increase, it can still be considered price gouging. In a competitive market, prices are based on what the market can bear, not on moral considerations. In such markets, prices must remain low to remain in business. However, high prices without justifiable reasons could be indication of monopolistic behavior from too much consolidation, and should be called out and investigated.

Usually you don't consider health insurance to be a moral hazard in this way. However, in healthcare, a moral hazard can occur when people use more healthcare services because the cost for them is low.

What is the role of an engineering manager exactly? I've been thinking about this question for awhile and I get wildly different answers depending on who I ask. It seems to me that an engineering manager contributes to the intellectual product, attempts to solve communication issues between teams, and fills in gaps for what the team needs. It is amazing to me that there is so much variance for something that should be a well defined role.

When a company takes an action solely to limit competition with no consumer benefit, they should be immediately investigated for antitrust behavior.

The more people you survey, the higher the confidence that the questionnaire is representative of the entire population, or the 'confidence interval'. Since the population was randomly sampled & non-responses don't correlate to the topic questions, you can assume that the 90 students are representative of the entire population with +- some confidence interval.

The topic for this questionnaire is "student life, academics, athletics", which is an important part of the question when determining if non-responses are important. In this case ChatGPT gets it correct, but explains the reason very poorly.

Given the topic is "student life, academics, and athletics", would you care about the non-responses? If the topics correlate well to non-responses then you'd have a reason to care about the missing results. The best answer here is D, because there is always attrition when doing surveys and getting 100% responses isn't usually a goal for this type of experiment.

There is often attrition when doing surveys. If the correct answer is always A then that implies there can be no attrition for completing the survey. Since the list of topics asked don't seem to correlate with non-responses (such as doing a survey about sexual harassment where non-responses may be very important), the correct answer would be likely D, to discard the subpopulation.

Twitter's actions seems like retaliation against the employee for enabling other employees to collect evidence of discrimination. Performance reviews often exist to reduce the companies liability when firing people off, but being fired with good performance reviews can work against them.

What they did was collude to control supply of lessors by setting rates so there would be less competition. This is the definition of collusion and almost textbook anti-trust. The question for anti-trust cases like this is what was their enforcement mechanism.

The consumer market ban happens when the FCC refuses to certify any new devices so they are not eligible for import for the US consumer market.