Given Google's resources, I'm a little surprised they having created an LLM that would rewrite Chromium into Go/Rust and replace all the stale libraries.
MIT have developed a technique called Self-Adapting Language Models (SEAL), which enables LLMs to continuously improve by generating their own synthetic training data and updating their internal parameters in response to new information.
ToolAlpaca, InterCode and Reflexion are taking different approaches among others.
Animals think but come with instincts which breaks the output relative to the input test you propose. Behaviors are essentially pre-programmed input from millions of years of evolution, stored in the DNA/neurology. The learning thus typically associative and domain-specific, not abstract extrapolation.
A crow bending a piece of wire into a hook to retrieve food demonstrates a novel solution extrapolated from minimal, non-instinctive, environmental input. This kind of zero-shot problem-solving aligns better with your definition of intelligence.
There are levels to being a dick. I think that chronically online types tend to forget that at the other side of the screen there are real flesh-and-bone people who would find it unacceptable to be addressed in a disrespectful way.
According to your numbers, Moderna got lucky at a 10% chance of producing the mRNA COVID-19 vaccine in 48 hours of computation? I don't know, but there seems to be more factors at play.
Moderna got lucky in that we know enough about that virus that the chance of a COVID-19 vaccine was a lot more than 10%. The more general case of a drug is a lot more than specific one
> Also, in 2018, just one year before the outbreak in a DARPA grant proposal, Wuhan Institute of Virology and its collaborators proposed to construct genetically modified SARS viruses having a furin cleavage site, a feature associated with increased viral growth and increased transmissibility. They proposed to insert the furin cleavage site at the spike gene S1/S2 border, and to construct the viruses by synthesizing six nucleic acid-building blocks, and assembling them using the reagent BsmB1.
> Fourth, in 2019, a novel SARS virus having a spike with extremely high binding affinity for human SARS receptors, a furin cleavage site inserted at the spike S1/S2 border, and a genome sequence with features enabling assembly from six synthetic nucleic acid building blocks using the reagent BsmB1--a virus having the exact features proposed into 2018 NIH and DARPA proposals--emerged on the doorstep of Wuhan Institute of Virology.
> Taken together, the presence of a spike having an extremely high affinity for human SARS receptors, the presence of a furin cleavage site inserted at the spike S1/S2 border, the genome sequence enabling assembly from six synthetic nucleic acid- building blocks using the reagent BsmB1, and the one-for-one match between these features and the features proposed in the 2018 NIH and DARPA proposals, make an extremely strong case--a smoking gun--for a research origin.
It raises questions that James Comer keeps raising to attack Fauci, but which never seem to get anywhere close to being proven. It's now been four years. The various libels against Fauci remain unsupported.
The closest I got to Linux mobile is GPD Pocket 4 with LTE and regular apps. Since I can get it to cap at 5 watts, it can give 9 hours of battery life. It does most things I care about, but it is just a mini laptop (which is good enough for me).
I remember embedding assembly inside my Turbo Pascal procedures and functions, using asm/end blocks. It worked quite well. External assembly module object linking was an option too.
It did in Object Pascal as well - way back in a different life when I hung around a programming channel there was a big argument about C always been faster than Pascal from one of the C devs in there - so we set the challenge as "count instances of several words in a 100MB file" fastest using only language features and the standard library for each (what came out the box).
I beat the C programmer because I "cheated" - I dropped into assembly for parts of it since that was part of the core language to be able to do that, he used whatever string library was shipped with his compiler at the time, he rigged it because Object Pascals standard string functions where notoriously slow (fast enough but much slower than C implementations of the time) - one of my prouder moments and 100% not something I could do now, I've completely forgotten assembly.
There is no reason heroism cannot be a sustainable business model. Social entrepreneurship, B-corps, even Patagonia or Ben & Jerry's embed environmental or social activism directly into their core strategy and values.
That LLMs have some basic metaknowledge and metacognitive skills that they can use to reduce the hallucination rate.
Which is what humans do too - it's not magic. Humans just get more metacognitive juice for free. Resulting in a hallucination rate significantly lower than that of LLMs, but significantly higher than zero.
Now, having the skills you need to avoid hallucinations is good, even if they're weak and basic skills. But is an LLM willing to actually put them to use?
OpenAI cooked o3 with reckless RL using hallucination-unaware reward calculation - which punished reluctance to answer and rewarded overconfident guesses. And their benchmark suite didn't catch it, because the benchmarks were hallucination-unaware too.
reply