sota_pop
e.g. “Pat” is cited, but doesn’t specify whether that refers to Patricia or Patrick.
Only having experience within the US myself, I have always found the idea of “forced ranking” for grades a little odd. Though admittedly, I do skew towards being (perhaps a bit too) idealistic in that sense.
How is priming different from transfer learning?
Grats to the team.
How is this different from nearest-neighbor within traditional vector search, used by RAG systems? What am I missing?
1. Compute input embeddings 2. Concurrently compute distance-to/logits-of runtime-specified (sic cached) label embeddings 3. Return greatest/closest label
I do appreciate the Kahneman reference.
MacOS certainly feels like a toy to me in many regards.
60% of the time, the framework works, every time.
The initial debacle that was the Recall recall and announcing the level of AI stuffing of win 11 that made me actively try out Linux mint (which has been quite nice and smooth). LLMs help to diagnose and fix items if you aren’t familiar with Linux. You no longer have to scour message boards for up-front answers (only use them as a “confirmation backstop”).
For work I now use macOS which feels like a toy but is ok overall (probably still some personal habit-based things).
If it weren’t for the fact I still develop plugins for a single, very specific windows-only app in my spare time, I would completely ditch windows, with contempt. Outside of Windows I am continually stubbing my toe on every single use of every single MSFT product with which I interact; primarily outlook, teams, and some azure stuff. Their products feel highly “feature optimized” with absolutely zero (perhaps negative) cross-feature cohesion. e.g. it’s as if the various outlook panes and the boxes of content were designed so modularly and independently that they don’t re-size or scale themselves cooperatively.
I perpetually experience stark jamais vu using their products, i.e. I am forced to use them every day and each time feels as unfamiliar, clunky, and unintuitive as the first time I ever touched them.
THEY’RE USING YOUR DATA
As they involve themselves directly, explicitly in the process of scientific research and discovery, it becomes a question of scientific and professional ethics.
Strange times.
- LoZelda Ocarina of time& majora’s mask
- starfox 64
- Mario party
- extreme G racing
- jet force Gemini
- golden eye
- conkers bad fur day
- Mario 64
- Tony hawk
Man… so much nostalgia
I mostly used KA between 2010-2015, and received immense benefit. These were the years it was MOSTLY “just videos”. Having revisited more recently, I found the gamification, quizzes, etc. simply distracting.
My understanding is that he started KA as a private YouTube channel to tutor his own family members, made it public, and it grew in popularity organically.
I haven’t used khanmigo, but given the climate and evolution of LLMs over the last 6 years, it is a completely reasonable goal to attempt to employ them in learning (though it sounds like it was not as effective as planned for.. reasons).
KA never pretended to be “all you need” for learning. It was always at least an excellent supplement to a given curriculum, and Sal’s genuine interest in the videos he narrates always sounded (to me) as palpably genuine which also goes a long way. There is a reason most textbook chapters have an initial section called “motivation”. The primary job of a teacher is to incite interest, and potentially direction at strategic points…
As the adage goes: You can lead a horse to water, but you can’t make it drink.
The names just hit different.
The article was a nice read, but I was having a hard time understanding why attempt to coin a new word when the exact phenomenon is already (very very) well known.
Between the years of 2018-2025, I led very many professional discussions and educational presentations regarding the theory and application of AI/ML in the org I worked (a ~120 person org lacking in any substantive expertise or awareness of the field).
One of the most important topics within those conversations was demystification. Several times I was asked (in earnest, by adults, in public fora) about “terminator/skynet scenarios”, sometimes intended metaphorically (related to the profession) and sometimes literally.
My response every single time:
I don’t fear “skynet”. My biggest fears surrounding the continued proliferation of AI are the mundane, boring, and nascent uses that completely bastardize the world as we know it. Uses like mortgage/loan approval, healthcare coverage approval, hiring, legal adjudications etc., etc.
Uses where the decision outcomes have a direct, major, intimate impact on people’s lives.
I am generally pro-AI/ML (when applied thoughtfully), but my greatest fears are and have ALWAYS been in knowing just how much easier it enables these kinds of decisions to inevitably be outsourced or made in an otherwise (more) thoughtless, dehumanizing way.
While it’s easy (and at times cathartic) to criticize big corp leaders, the CEO and company described in the article gives me a visceral reaction deeply within the uncanny valley that is actually difficult to describe.
sometimes starting from scratch just faster and/or easier.
- any project that involves hardware or physical objects (especially ones you are manufacturing!) is impressive
- the craftsmanship
- references to established standards
- documentation
- actually publishing
As others have said here you can drop the “wannabe” tag (or don’t, we’re all wannabes at heart :))
Nice job and keep up the good work!
I would have liked a comparison to something like Kanban which tries to aggregate to a (seemingly?) similar metric but is a more prominently known system.
With interest in this space you may be interested to check out:
- That Open Company in the BIM industry (not personally affiliated, but admirer). - Fovea ArchVision (also not directly affiliated, but can attest that the team are great people.)
It is _intelligent_ in the sense that it optimizes a thing that is hard for humans to not anthropomorphize.
Things like percolation theory, swarm theory, and others are similar topics in “complex systems”. Neural networks are interesting because they combine aspects of both complex systems and dynamical/adaptive systems (ie systems with a feedback loop).
A neural network at its very core is a function fitting algorithm. It stores matrices of parameters (ie weights and biases) such that every parameter (and combinations thereof) captures a relationship of your data in exactly the same way the slope and intercept are obtained through Linear Interpolation. All of the Regularization tricks applied just are attempts to incentivize a given parameter to not encode any trends too specifically (think an instance vs a type).
In this way you can ask yourself “how many aspects of your data are required to capture it adequately?” This is what scaling offers.
Contributors responded by going to delete their own respective contributions en masse. Upon doing so, were banned by the platform mid-process which then led to people going back to revise their contributions to be false rather than deleted.
I guess that’s “LLM related”