For two consecutive quarters, generative #AI dealmaking at the earliest stages has declined, dropping 76% from its peak in Q3 2023 as wary investors sit back and reassess following the initial flurry of capital into the space.
If you're curious, the bot just asks a Mistral Instruct model to come up with a prompt based on a random line from /usr/share/dict/words, then passes the response to stable diffusion.
Seeding it with a random word for inspiration helps the model come up with different stuff. Without some random input, the Mistral tended to describe the same or similar prompts every time. I don't blame it, it's only the baby 7B model. It is doing its best!
I am currently the only individual in my organization who has a Microsoft 365 Copilot license.
In other words, I am the pilot for the Copilot pilot program.
So far, I have not seen any set-the-world-on-fire (metaphorically) capabilities for this set-the-world-on-fire (literally) technology. Having Copilot summarize long email threads is pretty good. But its much-touted ability to synthesize data from across the Microsoft tenant is, so far, not impressive.
#LLM bawi i uczy!
Jednak ta cała tzw. sztuczna inteligencja nie jest taka zła jak ją malują. Dzisiaj dzięki jej pomocy, nauczyłem się korzystać z API #Mastodon do wyświetlania określonych wpisów w przeglądarce 😁
Jestem o krok od stworzenia własnego klienta!
“I speak to a lot of businesses around #AI, and particularly #GenAI, and I’m sensing a #hype fatigue. Part of this is due to the challenging of bridging the gap from PoC to production"
the energy cost of training an #LLM is about the same as the energy required to raise 2 kids. But unlike kids, you can copy #LLMs and amortize that energy cost across inferences
"Reading, understanding, and fixing code written by others consumes 90+% of the time a programmer spends ..."
The complete blog post is somewhat on the long side, but the key point it makes is worth repeating, often, as it has been demonstrably true since the 1950s and is highly likely to be true for the next 70 years.
I have been using brave search for the past week.
It has a chatGPT kind of thing, but when the #LLM doesn't know, it says it doesn't know, and when it claims to know, it shows where it got the information. Its answers are not always correct, but by looking at its sources, that's possible to see.
So Microsoft is shipping a new feature in Windows 11 called Recall, which takes screenshots of what the user is doing every few seconds, and then feeds it into OCR.
And I've read a number of people describe it as useless.
But I disagree.
I'm sure plenty of people will find it very handy.
For example, your friendly local law enforcement agents and prosecutors are likely to find a feature like this very useful.
As will the NSA, the other three-letter agencies in the US, and intelligence agencies around the world.
Including the ones in authoritarian states. A couple of back doors, and it will be so much easier keeping track of who's been typing naughty words like "Prigozhin", "Navalny", or "free Hong Kong".
Not in the state surveillance business? No worries!
Assuming this data isn't locked down properly — and we are talking about Microsoft here — it's sure to find plenty of more mundane uses.
Perhaps for bosses who will no longer need to install keyloggers to snoop on their staff.
Or jealous current and former partners.
Mark my words, this poorly-thought-through attempt to shove LLMs in another place they don't belong to temporarily spike Microsoft's share price will find its uses.
And the next computer I get definitely won't be running Windows.
This post by @maggie has some great ideas on how #LLM tech can help enable #LocalFirst applications for regular folks. I've been wanting to do something similar within @agregore some day with local LLMs helping people author p2p web apps.
1/3 I tested some popular latest LLM UIs for accessibility with screen readers, including oobabooga text-generation-webui, Open WebUI (aka Ollama WebUI), GPT4All, LM Studio, Koboldcpp, and Llama.cpp server on Windows. The most accessible was Llama.cpp server, though it had the fewest features. Oobabooga was also good, except for the list box not announcing choices as you browse; however, you can check your selection afterward. #accessibility#LLM#AI
and there are so fucking many things wrong with it
one of the most amazingly wrong things is that... they're already throwing "ai" bullshit at these screencaps they're doing every five seconds, right? that's what does the OCR and also does the LLM-driven description for the search functionality later
and yet no one
NO. ONE.
thought to tell it
"and don't save screens with the word 'password' on them."
YOU COULD DO THIS WITH GREP, YOU STUPID FUCKS, WHAT THE HELL IS WRONG WITH YOU?! IT'S NOT HARD!
Pünktlich zur #bibliocon24 starten wir im VÖBB einen neuen, experimentellen Dienst: den VÖBB-Chatbot. Als meines Wissens erste (?) deutsche Bibliothek kombinieren wir hier Sprachtalent und "Wissen" eines Large Language Models (#LLM) mit den vollständigen Metadaten unseres #VÖBB Kataloges (als sog. Embedding).
With #LLM applications more abundant, have researchers been using them to assist their writing? We know they have when writing peer reviews [1], but how about doing so in writing their published papers?
Liang et al comes back to answer this question in [3]. They applied the same corpus-based methodology proposed in [2] on 950k papers published between 2020 to 2024, and the answer is a resounding YES, esp. in CS (up to 17.5%) (screenshot 1).
I made good progress on the #AI collaboration project with @Jorvon_Moss over the weekend. The #Nvidia Orin Nano boots up the servers and WiFi hotspot automatically. You just need to run Hopper Chat on the #RaspberryPi. No internet required! #LLM#ChatGPT
I have a newly graduated SW Eng (BS in CS) who is struggling to find a job and getting advice to go back and get a Master’s Degree in #LLM in order to be more marketable.
I’ve always heard that grad degrees aren’t strictly necessary in SWE to start but is this changing? Are there other time investments that make more sense (open source contributions, certifications, personal projects, etc?)?
Weekend discovery. An intermediate step in the RAG process is document chunking. Determining the appropriate chunk size can become a trial & error game. James Briggs does a great job of explaining how to use Semantic Chunking to get better results.