jmalicki

5 ms·

61.
by jmalicki 2mo ago
That distillation exists isn't the question. It's about evidence this is an active force in competition in LLMs.
62.
by jmalicki 2mo ago
There is GPT 5.6 Sol on Cerebras if you want to try that experience for an ungodly sum of money (not getting into GPT 5.6 Sol vs Fable, but only one is available on Cerebras) for an 11x speedup.
63.
by jmalicki 2mo ago
AI is pretty universally hated, it might be the one thing that large swaths of society have unity about! It's right up there with thinking babies and puppies are cute.
64.
by jmalicki 2mo ago
> At some point the size of the project becomes too big for LLM to fit in its context window, and with the tendence to add code rather than remove, the bloat can explode really quickly. While it's certainly worse with LLMs, humans,
65.
by jmalicki 2mo ago
Because oil, cigs, and beer don't create the existential threat to the human race that Anthropic is trying to claim it creates. Yes, it is software , something more powerful and thus potentially more harmful than any of the above phys
66.
by jmalicki 2mo ago
Uber called out that they were aggressively doing an end run around both taxi licensing AND minimum wage laws and facing significant pushback on both fronts, among many other risks.
67.
by jmalicki 2mo ago
"Our AI products and X platform are subject to complex and evolving U.S. and foreign laws and regulations regarding privacy, cybersecurity, data use, data combination, data protection, content, AI, competition, youth protection, safety
68.
by jmalicki 2mo ago
Uber and Airbnb both had similar risks in their S1s, for obvious reasons.
69.
by jmalicki 2mo ago
They can probably hire them, and they'd get PIP'd for not jumping when management says jump. Toxic corporate culture can go a long way to self-sabotage.
70.
by jmalicki 2mo ago
Dan Luu might benefit from having an LLM optimize his website for readability.
71.
by jmalicki 2mo ago
AI labs are already working on getting LLMs to control nuclear reactors, and hiring experienced reactor operators to help train them: https://www.reddit.com/r/nuclear/comments/1qnzxvg/this_job_p...
72.
by jmalicki 2mo ago
Except they all do, and the money flows to Cerebras :)
73.
by jmalicki 2mo ago
It's not like there are only two buckets: 1. HFT doing ass-simple arbitrage where only latency matters 2. More sophisticated slower trading taking in deeper signals Those are two points along a continuum. If you are reacting to an ear
74.
by jmalicki 2mo ago
When you FOIA body cams that will often redact the faces of other individuals. Likewise I would expect that the license plates read be redacted. It makes sense to protect it since you are exposing members of the public, not the government.
75.
by jmalicki 2mo ago
And do! A lot of newer home security cameras have ALPRs so you can set up rules like "if a car with a plate number not on this whitelist pulls up on my driveway send me an SMS"
76.
by jmalicki 2mo ago
What else are you going to write CUDA kernels in that will be better than Mojo? This doesn't compete with Python. Like if someone wrote a competitor to VHDL that looked like TEMU Python... Python would not be its competitor, VHDL would
77.
by jmalicki 2mo ago
> Eventually people will start adopting the distribution used by LLMs, making classification harder. I recently heard someone say "that's genuinely the exact solution I was looking for" and had to do a double take.
78.
by jmalicki 2mo ago
For these sort of things, I wish they would have some semi-stable beta or prerelease tracks other than just nightly so you could use the new features on a somewhat stable branch. If something is in nightly that is too crazy for a lot of pr
79.
by jmalicki 2mo ago
We were talking about examples of compiler optimizations that save memory, so I gave one. Value classes are not a compiler optimization, and I wasn't talking about the standard, I was talking about an implementation.
80.
by jmalicki 2mo ago
pytorch is really mostly C++ after all, you don't have to use it with python. Python just allows you to use it like "normal code", whereas if you program it in C++ you really get exposed to what's under the hood and can&
81.
by jmalicki 2mo ago
Does 1 really happen? I would never trust a compiler where 1 was a possibility. If it can work it should.
82.
by jmalicki 2mo ago
JVM does a lot of escape analysis to turn heap allocated memory into stack local variables. It doesn't matter if it's local since it's a VM, it's doing it at runtime and can change an entire call stack of non local code
83.
by jmalicki 2mo ago
Who other than Noctua would notice?
84.
by jmalicki 2mo ago
"What kind of stocks should my mutual fund invest in?" "What sort of startups should I fund?" "Should my business aggressively hire this quarter anticipating growth, or hold out for a possible recession?"
85.
by jmalicki 2mo ago
This is very common for economic statistics. GDP too. That's how you can get more accurate information, and you have actual numbers on the discrepencies. They rely on the household data more because it doesn't have as much syste
86.
by jmalicki 2mo ago
The main text doesn't plot the error bars, and the full data tables in their release don't have error bars either. You have to really dig into their website to find their std errors.
87.
by jmalicki 2mo ago
No, they shouldn't. Timely estimates as to the best knowledge of our current situation is extremely important for decision making under uncertainty. If the error bars are large, that is important information to know.
88.
by jmalicki 2mo ago
Not in their primary news release. A footnote says it is "generally conducted at the 90% interval", but the individual numbers do not have their actual computed error bars. Yes, they do publish them, but it is disappointing their
89.
by jmalicki 2mo ago
DOGE hurt them pretty hard: "The Bureau of Labor Statistics posted a notice on Wednesday stating that it stopped collecting data in three not-so-small cities (Lincoln, Nebraska; Buffalo, New York; and Provo, Utah) and increased “imputa
90.
by jmalicki 2mo ago
That's true for the establishment survey, but the population survey, where they contact individuals, is what is actually used to calculate these numbers. It is literally calling/writing 60,000 people and asking "are you emplo

More ›