@tamaybes Given the jump wouldn’t we expect it to solve at least one currently unsolved research question. If not why?
English
Peter Björkmarker
693 posts
1/11 I’m genuinely impressed by OpenAI’s 25.2% Pass@1 performance on FrontierMath—this marks a major leap from prior results and arrives about a year ahead of my median expectations.
English
@mmjukic Agreed. My 2 cents. You will always need to give an LLM a quite large and specific context for how to operate. This makes them much more useful than you might think but also much less impressive.
English
Don't tell me the AI is going to replace all human beings thanks to its unquestionable superiority but then get mad when I judge the AI like I would a human being.
English
I don't understand why everyone pretends like the unreliability of LLMs when it comes to facts and reality is some minor flaw rather than a complete dealbreaker.
An intern-on-demand is already pretty useless. One who randomly lies too gets the lash and the boot, not the job!
rawn@reallyrawn
dude FUCK chatgpt. i spent like the past hour working out a plan to host my haskell project on cloudflare workers (chatGPT recommended). then guess what? cloudflare workers don't support haskell at all. wtf was the plan?
English
@perrymetzger Yeah I can attest to it being challenging on multiple levels…. What would you suggest we do? Given we want to build rather than sit in meetings…
English
And especially this bit.
x.com/itsolelehmann/…
Ole Lehmann@itsolelehmann
Europe chose security over growth. America chose innovation over regulation. The results? America has produced 9 trillion-dollar companies (9/10 of the most valuable companies in the world). Europe? ZERO. Nowhere to be found:
English
None of this matters. There will be no AI companies in Europe, and in the medium term, companies from other countries will just write off the EU and not sell products to them any more either because it won't be worth it given the decreasing size of their economy.
Luca Bertuzzi@BertuzLuca
🚨NEW: The European Commission has just published the first draft of the Code of Practice for general-purpose #AI model providers. digital-strategy.ec.europa.eu/en/library/fir…
English
@JagersbergKnut Hope you are correct! However, given LLMs ability to also produce vast lengths of text without actual information it might also end up being the opposite. en.m.wikipedia.org/wiki/Jevons_pa…
English
RAG LLMs have practically obsoleted the human capability of skim reading .
this is so great, because it will economically obsolete the abundant seo slop that perfectly serves skim reading without providing any actual information.
attention will go more where it belongs.
English
@simonw @emollick Agree! OpenAI will have trouble getting enough context to do that. Regardless of superintelligent capability “perfect” does not contain much information. A privacy focused Apple could potentially know a lot more about what “perfect” means to you. Context wins in the long run.
English
@emollick I'm much more convinced by the Apple approach than the OpenAI one - the whole "book me the perfect restaurant / holiday based on my known preferences" thing feels like an unrealistic fantasy to me, no matter how good the models get
English
Apple is betting on AIs based around lightweight reasoning and very good local retrieval/ low hallucination rates “what’s my dinner plans with mom?”
OpenAI is betting on autonomous agents: “Set up a meal with my mom somewhere she likes, and maybe a show afterwards, in my budget”
English
@simonw We use it to take lazy/ambiguous user questions, add more specific context and ask the llm output a more “better” question. We feed this question to a texttosql llm.
English
What are the simplest useful examples you've seen of prompt chaining, where the output of one LLM call is used as input to another?
I have seen plenty of theoretical examples, but I'm looking for some concrete results
English
Peter Björkmarker retweetledi
There will be no edge in talent in AI
There will be no edge in compute in AI
There will be no edge in models in AI
Will be
Data
Distribution
Integration
Google et al will drive generalised AI to zero marginal cost & we will make open variants of cutting edge open & available
Emad@EMostaque
I have no idea what you would spend that much money on Reality is all claims that you’ll need to spend hundreds of millions and billions training models are dumb and largely to raise money and scare folk off Will see rapid commoditisation in this space in the next few years
English
Peter Björkmarker retweetledi
A big change for me in programming in GPT-4 is I now don't factor in the overhead of using an unfamilar language or library in spiking out solutions to most problems. For rapid prototyping, exploration, and throwaway scripting, this is an absolutely insane change.
English
Struggling with some simple real-world problems even with GPT-4
While AI researchers are working on aligning some imaginary superintelligence 🤔
To date, AI hasn't been capable of replacing even a single profession reliably.
There is a huge disconnect between research and the real world.
Also, no one has scientifically proven that autoregressive LLMs trained on human-generated data can somehow be capable of super-human intelligence. Logically speaking, their ceiling should be human intelligence 🙄
Let's please get to human-level reasoning reliably first. 🙏🙏
English
Peter Björkmarker retweetledi
I suspect that there's no explanatory theory linking those proxies to anything that would reasonably be regarded as competence in practical applications.
No doubt slide-rule test scores have decreased too.
What causes moral panics? I don't know. Not iPhones.
Matthew Yglesias@mattyglesias
Very broad nature of the decline makes me suspect "the phones"
English
Peter Björkmarker retweetledi
I don't even know if you could design software non-iteratively today. Almost all the issues I run into these days is things like "I guess this vendor has an undocumented rate limit so we'll have to refactor and make requests in batches" and how do you plan for that?
English
@ddlabar The way to test it. Skip explicit dispose. Put the API under a bit of load and you will get very strange behavior and exceptions.
English
@ddlabar I have not been able to reuse the same container consistently across multiple plugin executions. I do not think you can as plugins need to be stateless.
English
I had a conversation online somewhere with someone that suggested a dependency injection library that is a single file that might work inside of #Dataverse plugins but I can't seem to find it, or more importantly the suggested project. Anyone help?
English
@ddlabar Note that you have to instantiate and explicitly dispose the container in the execute method as the platform does not always run the plugin constructor (happens often under a bit of load)
English
@ddlabar New try!
github.com/grumpydev/Tiny… is what I suggested to you the last time around I think?
Works well and allows for simplifying the code base into small services as needed as lifetime management.
….
English
🌟Thrilled to join Microsoft as a Cloud Advocate 🥑
Exciting times ahead working with @aprildunnam and @laskewitz. 🙌Grateful for all your ongoing support and encouragement. I'll be just as active in community, events, & supporting Ribbon Workbench!! 👨💻🎉 youtu.be/tqq4kDTVwBg
YouTube
English
@benedictevans Surprised this pattern isn't useful for you: twitter.com/emollick/statu…
Ethan Mollick@emollick
I gave Claude the descriptions of 400 of the latest Y Combinator startups & asked it to generate on-trend startup concepts based on them. Some interesting ideas. (As a general principle, this type of few-shot prompting, where you give the AI examples, is an effective technique)
English