jmalicki

14 ms·

361.
by jmalicki 6mo ago
What are good harnesses? I haven't yet been able to get good agent teaming approaches out of other harnesses yet, before that feature I mostly regarded the space as competitive, but until another harness can do as well with Claude mod
362.
by jmalicki 6mo ago
It's a great way to get funded by your CEO and get good performance reviews; xAI employees know how their bread is buttered.
363.
by jmalicki 6mo ago
The soft max is the probability of the next token being whatever in the training data conditioned on the inputs. The author just doesn't know that apparently and thinks it was an arbitrary choice. The author's essay on the sigmoi
364.
by jmalicki 6mo ago
Hogwash. If you get into deriving maximum entropy distributions via the calculus of variations, the multinomial is the maximum entropy distribution among categorical distributions. This is exactly the sense that it comes up for old school
365.
by jmalicki 6mo ago
But the Mac Studio can run LLMs reasonably faster better than non-enterprise-server setup. That is the only thing that matters at this point - it's an LLM accelerator, not a personal computer, at this point.
366.
by jmalicki 6mo ago
How a website is going to identify your PI as non-fake? Isn't that the entire business model behind Persona which has been in the news for leaks? (There are a few websites I've had to verify my if I with them for)
367.
by jmalicki 6mo ago
Meta is a huge competitor. The have less fewer ads for directed searches of course, but know you were browsing products on the web so can spam you all over for the types of things you were shopping for.
368.
by jmalicki 6mo ago
Of course that can happen! But they are actual tokens produced, that are then read by the answer generation as part of the prompt, nonetheless. And the hidden state of course has a ton of logic that may not be apparent by the tokens produc
369.
by jmalicki 6mo ago
Have you seen documentation that the thoughts in Claude Code are slipped out separately, authoritative or otherwise? I've heard this claimed a few times and wondering what they're doing differently from traditional thinking model
370.
by jmalicki 6mo ago
Same for SciPy (At least the last time I dove into it around 10 years ago). A lot of the C code you see for numerics is a straight up f2c run checked in.
371.
by jmalicki 6mo ago
SQS is particularly hard, there are other options, but it has less direct competitors than S3
372.
by jmalicki 6mo ago
It's very much necessary but not sufficient. In real life the sample complexity matters a lot too, which is also asymptotics, but a more important one. E.g. how the central limit theorem is far more powerful than the law of large num
373.
by jmalicki 6mo ago
Yes, you do linear interpolation between an infinite number of data points.
374.
by jmalicki 6mo ago
Data structures are not tactical components of programming languages. E.g. when I am writing SQL, I need to be thinking about the underlying data structures too - even though I am not specifying the execution path.
375.
by jmalicki 6mo ago
To some degree IMO big data is still a mindset when it might take a day to process your data in a normal SQL query. Some tech doesn't scale to the data size for all use cases, and you need different solutions.
376.
by jmalicki 6mo ago
Totally! I mean, the same could be said of painstakingly hand coding assembly language - that today's developers haven't done so is what leads us to bloated electron apps, so there is something lost! But the larger scale system d
377.
by jmalicki 6mo ago
> How do you think engineers in the second half got there? By writing tons and tons of code to "build those reps" and gain that experience. It's not by writing syntax that you get there. It's by creating software and
378.
by jmalicki 6mo ago
> Sure, there's a very gradual, strictly limited, tightly controlled rollout By city and ODD, but if you are in the area and your source and destination for the route are both within the ODD they are just as available as an Uber.
379.
by jmalicki 6mo ago
At the very least you should give it a non-prod copy of the database, not direct access to the DB actively powering production right now. I've done work for a hedge fund where the DB ran directly on the manager's desktop. I worke
380.
by jmalicki 6mo ago
Its output quite literally is not independent, as the "thinking tokens" are attended to by the attention mechanism.
381.
by jmalicki 6mo ago
If the cyclist was doored by an exiting passenger, would t that imply it should further block the bike lane to increase safety as it is not safe for a bike to pass while a passenger is exiting? If the car door opening is what injuries the
382.
by jmalicki 6mo ago
So it is equal to what neuroscientists and psychologists have proven about human beings!
383.
by jmalicki 6mo ago
It does have access to its thoughts. This is literally what thinking models do. They write out thoughts to a scratch pad (which you can see!) and use that as part of the prompt.
384.
by jmalicki 6mo ago
Luckily since I met this guy named Claude most of that complexity has gone away.
385.
by jmalicki 6mo ago
Part of the issue with neural nets is that historically they were next to impossible to train. ADAM, BatchNorm/LayerNorm, initialization schemes, and GPUs for pure speed really helped to change all of that.
386.
by jmalicki 6mo ago
Part of the issue they mention is contamination - the tests are in the training data. The other issue they mention is being overly constrained vs. what is asked for - such as requiring specific class or function names to pass that were not
387.
by jmalicki 6mo ago
Has it been reasonably possible to overfit to the errors in ImageNet, or are they effectively random noise?
388.
by jmalicki 6mo ago
Most places I've worked have roadmaps, i.e. investable priorities. If you can burn through lower priority experiments quickly it's great! They might be working on all of the super high level things they can think of, but there are
389.
by jmalicki 6mo ago
While value isn't linear, prejudgement of value for allocation of resources is very imperfect. A lot of stuff that doesn't make the cut is the the stuff that does have value. When you're lowering the bar, remember it's
390.
by jmalicki 6mo ago
The properties that the uniform approximation theorem proves are not unique to neural networks. Any models using an infinite dimensional Hilbert space, such as SVMs with RBF or polynomial kernels, Gaussian process regression, gradient boost

More ›