logsr

The biggest problem OAI has is that they don't own a data source. Meta, Google, and X all have existing platforms for sourcing real time data at global scale. OAI has ChatGPT, which gives them some unique data, but it is tiny and very limited compared to what their competitors have.

LLMs trained on open data will regress because there is too much LLM generated slop polluting the corpus now. In order for models to improve and adapt to current events they need fresh human created data, which requires a mechanism to separate human from AI content, which requires owning a platform where content is created, so that you can deploy surveillance tools to correctly identify human created content.

Love this. I have been looking for an app that does this. Congrats on the launch.

In a few more years there will probably be virtually no human users of web sites and apps. Everything will be through an AI agent mediation layer. Building better CAPTCHAs is interesting technically, but it is doubling down on a failed solution that nobody actually wants. What is needed is an authentication layer that allows agents to act on behalf of registered users with economic incentives to control usage. CAPTCHA has always been an economic bar only, since they are easy to farm out to human solvers, and it is a very low bar. Having an agent API with usage charges is a much better solution because it compensates operators instead of wasting the cost of solving CAPTCHAs. Maybe this will finally be the era of micro payments?

effectiveness, professionally, is creating value for your clients and employers. the ratio of value to cost is how efficient you are as an engineer. it is all efficiency.

Just going off specs sheets from manufacturers and reviews (mostly consumer products, so enterprise should be the same or better).

There are only a few major NAND manufacturers: Samsung, Micron, Kioxia / Western Digital, SK Hynix, and their branded products are usually the best.

There are also several 3rd party controller developers: Phison, Marvell, Silicon Motion, which I think are the largest, and then a bunch of others.

I hadn't looked at this in a couple years, so 16 channel controllers are more common now, but only on high end enterprise devices.

4KB random read/write specs are definitely not trustable without testing. They are usually at max queue depth and, at least for consumer devices, based on writing to a buffer in SLC mode, so they will be a lot lower once the buffer is exhausted. Enterprise specs might be more realistic but there isnt as much public testing data available.

> something unique and different about software engineering

how much does the strength/weight ratio of building materials improve every year? how much does the price/quality ratio of building materials fall every year?

software engineering is working with technology where compute capacity was doubling and price/performance was halving every two years for decades. this rate of change has slowed, but in a world where the price of everything else inflates, software is a rare field that works with a continually deflating capital base.

the uniqueness of software development is real and based on underlying physical/economic factors that exist in very few industries.

software developers are not unique. they are still human. but the profession is unique and the financial incentives are unique.

consistent, careful, workmanlike effort over time wins on average, but because the incentives are large, many people will take the risk to get exceptional rewards.

i approach software development like a professional athlete. i work on optimizing every aspect of my performance (physical, mental, social, technical) to be the very best i can be. anyone who applies this level of dedication in any field will be highly successful, but it requires dedication, sacrifice, and risk, which many people are not willing to accept.

the world is full of exceptional software and algorithms designed by unique individuals who developed things nobody else could or would. there is a reason why mathematical discoveries are named after the people who discovered them.

building software for large corporations necessarily revolves around median-level teams, because mean-reversion always happens at scale, but that is also why a great deal of technical innovation takes place outside of large corporations.

> Focus is a super power

this is crucial. from my own productivity I know that I can function at 1X or 10X depending on my focus.

being a great engineer requires practice most of all, and the consistency of focus during that practice will impact its value.

in my experience, engineering is all about efficiency, and as i have developed over time the scope of factors i take into account when calculating the efficiency of something has increased. in the beginning i only looked at the technical details of the implementation, and then over time that expanded to considering maintainability, team co-ordination, business objectives, etc.

the potential scope here is unlimited. when you start, just making something compile takes all of your focus, but over time as programming becomes reflexive you are able to expand the factors you take into account far beyond the immediate code, and it seems trivial by comparison.

It is hard to see how a court could decide that copyright does not apply to training LLMs without completely collapsing the entire legal structure for intellectual property.

Conceptually, AI basically zeros out existing IP, and makes the AI the only IP that has any value. It is hard to imagine large rights holders and courts accepting that.

The likely outcome is that courts rule against LLM creators/providers and they eventually have to settle on licensing fees with large corporate copyright holders similar to YouTube. Unlike YouTube though, this would open up LLM companies to class action lawsuits from the general public, and so it could be a much worse outcome for them.

Amazing presentation. It really helps to understand the concepts.

The only add is that it understates the impact of SSD parallelism. 8 Channel controllers are typical for high end devices and 4K random IOPS continue to scale with queue depth, but for an introduction the example is probably complex enough.

It is great to see PlanetScale moving in this direction and sharing the knowledge.

Large tech companies are inefficient and have very poor engineering productivity but this is largely for systemic reasons. They definitely do not have any mechanism for identifying high performers.

The basic problem they face is that they pay a fixed wage for complete IP ownership (horrible deal) and so they intentionally do not measure and reward actual performance because they do not want to compensate high performers based on the value they create.

There isn't really a solution because these companies are making rational profit maximizing decisions, it just happens that mediocrity and managed decline of locked in revenue streams is profit maximizing for them.

the "conflict-free" in crdt is like miller high life being the "champagne of beers." it really means conflicts ignored and conflicting items discarded by established rules, which works well for some use cases, but for many does not.

they point out that they are adding conditional append to a log, which gives the universality of a log data structure and a mechanism for implementing lock-free/wait-free concurrency algorithms. it is a complete toolkit for building distributed systems because you can build anything else on top of those primitives.

> log as a service

very exciting. this is the future. i am working on a very similar concept. every database is a log at its core, so the log, which is the highest performance part of the system, is buried behind many layers of much lower performing cruft. edge persistence with log-per-user application patterns opens up so many possibilities.