TheGhostRobot
@theghostrobot.bsky.social
Data analyst at MIT Sloan. Focused on pathways and operational tooling. Opinions are my own.
Reposted by TheGhostRobot
mastodon users being way more into the linux penguin than prehistoric megafauna should count as false advertising
December 28, 2024 at 2:17 PM
mastodon users being way more into the linux penguin than prehistoric megafauna should count as false advertising
Reposted by TheGhostRobot
AGI is going to prove a limited standard to think about because we will have Jagged AGI - superhuman at some tasks, weaker at others. Just because o3 is as good as the 175th best competitive coder on Earth (out of 600,000) doesn’t mean that it is as good as them at every task they do (for now)
December 20, 2024 at 7:05 PM
AGI is going to prove a limited standard to think about because we will have Jagged AGI - superhuman at some tasks, weaker at others. Just because o3 is as good as the 175th best competitive coder on Earth (out of 600,000) doesn’t mean that it is as good as them at every task they do (for now)
Well, I guess the "AI getting cheaper" is technically correct for price-per-token-compute?
But +$6k for running a 400 prompt test? lol
But +$6k for running a 400 prompt test? lol
December 20, 2024 at 11:18 PM
Well, I guess the "AI getting cheaper" is technically correct for price-per-token-compute?
But +$6k for running a 400 prompt test? lol
But +$6k for running a 400 prompt test? lol
I deeply appreciate Google letting the "chain of thought" process be visible in Gemini "Thinking"
December 20, 2024 at 1:14 AM
I deeply appreciate Google letting the "chain of thought" process be visible in Gemini "Thinking"
And yet another company deploys an in house version of langchain.
Branding this as 'reasoning' is... You know what, sure. I'm too tired to fight it today.
Branding this as 'reasoning' is... You know what, sure. I'm too tired to fight it today.
Google reveals AI ‘reasoning’ model that ‘explicitly shows its thoughts’
Google reveals AI ‘reasoning’ model that ‘explicitly shows its thoughts’
It sounds similar to OpenAI’s o1 reasoning model.
buff.ly
December 19, 2024 at 10:55 PM
And yet another company deploys an in house version of langchain.
Branding this as 'reasoning' is... You know what, sure. I'm too tired to fight it today.
Branding this as 'reasoning' is... You know what, sure. I'm too tired to fight it today.
I really need to follow more people on here because I'm inadvertently just lurking on like 3 writer's replies and it's starting to feel like an uncomfortable parasocial thing.
December 19, 2024 at 10:49 PM
I really need to follow more people on here because I'm inadvertently just lurking on like 3 writer's replies and it's starting to feel like an uncomfortable parasocial thing.
Finished Sapiens by Yuval Noah Harari.
It's fine. Not perfect, not awful. Could write a short essay on my issues with it, but it's not egregious or offensive.
That said, oh boy, I get why the "I don't read books" SF bro effective altruism crowd think it's A MASTERPIECE, and I hate it for that
It's fine. Not perfect, not awful. Could write a short essay on my issues with it, but it's not egregious or offensive.
That said, oh boy, I get why the "I don't read books" SF bro effective altruism crowd think it's A MASTERPIECE, and I hate it for that
December 19, 2024 at 10:18 PM
Finished Sapiens by Yuval Noah Harari.
It's fine. Not perfect, not awful. Could write a short essay on my issues with it, but it's not egregious or offensive.
That said, oh boy, I get why the "I don't read books" SF bro effective altruism crowd think it's A MASTERPIECE, and I hate it for that
It's fine. Not perfect, not awful. Could write a short essay on my issues with it, but it's not egregious or offensive.
That said, oh boy, I get why the "I don't read books" SF bro effective altruism crowd think it's A MASTERPIECE, and I hate it for that
For anyone who's fascinated by the innerworkings of grey-market internet: Lumen is a dizzying peek into just how overwhelming the sheer volume of DMCA filings are.
Millions of requests sent daily, obfuscated by shell identities. It's honestly a miracle anyone can ever backtrace bad actors.
Millions of requests sent daily, obfuscated by shell identities. It's honestly a miracle anyone can ever backtrace bad actors.
Luigi Mangione fan art, merch, and photos are getting hit with bogus copyright takedown requests all over the internet. Artists and journalists affected. Very difficult to tell who is doing it, but one company says they received a request from United Healthcare: www.404media.co/copyright-ab...
Copyright Abuse Is Getting Luigi Mangione Merch Removed From the Internet
Artists, merch sellers, and journalists making and posting Luigi media have become the targets of bogus DMCA claims.
www.404media.co
December 19, 2024 at 8:34 PM
For anyone who's fascinated by the innerworkings of grey-market internet: Lumen is a dizzying peek into just how overwhelming the sheer volume of DMCA filings are.
Millions of requests sent daily, obfuscated by shell identities. It's honestly a miracle anyone can ever backtrace bad actors.
Millions of requests sent daily, obfuscated by shell identities. It's honestly a miracle anyone can ever backtrace bad actors.
I haven't been keeping up with this the last few years, but man I've lost the thread on what streaming devices are doing in 2024
The Reckoning of Roku is such a downgrade from the past two Avatar novel series.
The way it handles Sozin and Roku's relationship is straight up terrible. Sozin himself is stripped of any redeeming qualities, his companions are boring clichés, and his role in the story feels almost shoehorned.
The way it handles Sozin and Roku's relationship is straight up terrible. Sozin himself is stripped of any redeeming qualities, his companions are boring clichés, and his role in the story feels almost shoehorned.
December 18, 2024 at 3:21 AM
I haven't been keeping up with this the last few years, but man I've lost the thread on what streaming devices are doing in 2024
Reminder the tuco tuco exists.
Its chromosomal diploid numbers range from 10 to 70, and is considered "taxonomic chaos"
They challenge our definition of species and make cute noises.
Its chromosomal diploid numbers range from 10 to 70, and is considered "taxonomic chaos"
They challenge our definition of species and make cute noises.
December 18, 2024 at 2:59 AM
Reminder the tuco tuco exists.
Its chromosomal diploid numbers range from 10 to 70, and is considered "taxonomic chaos"
They challenge our definition of species and make cute noises.
Its chromosomal diploid numbers range from 10 to 70, and is considered "taxonomic chaos"
They challenge our definition of species and make cute noises.
Tried to run "collection of agents" in a sandbox and the "project manager" bot started roleplaying and talking to itself eventually.
Even when all the subordinate functions went dark, it just playacted like it was getting updates.
Amusing, but not a great sign for a future of codebots
Even when all the subordinate functions went dark, it just playacted like it was getting updates.
Amusing, but not a great sign for a future of codebots
December 18, 2024 at 2:53 AM
Tried to run "collection of agents" in a sandbox and the "project manager" bot started roleplaying and talking to itself eventually.
Even when all the subordinate functions went dark, it just playacted like it was getting updates.
Amusing, but not a great sign for a future of codebots
Even when all the subordinate functions went dark, it just playacted like it was getting updates.
Amusing, but not a great sign for a future of codebots
Ok finally got them and they're worth it. Tru Tone nostalgia may be a generational thing but dear god I didn't know much I missed proper colored lights until I basked in their glow.
December 18, 2024 at 2:38 AM
Ok finally got them and they're worth it. Tru Tone nostalgia may be a generational thing but dear god I didn't know much I missed proper colored lights until I basked in their glow.
Gemini 2.0 is the only API I've found that doesn't "break character" for fact correction when it can use more in context ways of letting you know it disapproves.
Doubt this will ever be a widely acceptable way to test, but it's an impressive win for tooling
Doubt this will ever be a widely acceptable way to test, but it's an impressive win for tooling
December 18, 2024 at 2:16 AM
Gemini 2.0 is the only API I've found that doesn't "break character" for fact correction when it can use more in context ways of letting you know it disapproves.
Doubt this will ever be a widely acceptable way to test, but it's an impressive win for tooling
Doubt this will ever be a widely acceptable way to test, but it's an impressive win for tooling
Single monitor advocates are just outing themselves as not working in enterprise.
Yeah I too wish I could switch to an ultra wide, but have you ever tried opening an app as a remote desktop connection? Barely knows how to scale full screen
Yeah I too wish I could switch to an ultra wide, but have you ever tried opening an app as a remote desktop connection? Barely knows how to scale full screen
December 17, 2024 at 2:51 PM
Single monitor advocates are just outing themselves as not working in enterprise.
Yeah I too wish I could switch to an ultra wide, but have you ever tried opening an app as a remote desktop connection? Barely knows how to scale full screen
Yeah I too wish I could switch to an ultra wide, but have you ever tried opening an app as a remote desktop connection? Barely knows how to scale full screen
Mourning the loss of "it just works" tech stuff I can gift.
15 years ago, I could gift someone physical media an not worry about platforms.
Gadgets and one-off tech toys were less "what ecosystem" roulette. Could gift grandparents a Chromecast, or your aunt a cute Bluetooth speaker.
15 years ago, I could gift someone physical media an not worry about platforms.
Gadgets and one-off tech toys were less "what ecosystem" roulette. Could gift grandparents a Chromecast, or your aunt a cute Bluetooth speaker.
December 16, 2024 at 6:41 PM
Mourning the loss of "it just works" tech stuff I can gift.
15 years ago, I could gift someone physical media an not worry about platforms.
Gadgets and one-off tech toys were less "what ecosystem" roulette. Could gift grandparents a Chromecast, or your aunt a cute Bluetooth speaker.
15 years ago, I could gift someone physical media an not worry about platforms.
Gadgets and one-off tech toys were less "what ecosystem" roulette. Could gift grandparents a Chromecast, or your aunt a cute Bluetooth speaker.
me thinking I covered all my edge cases and forgetting about double top level country appended domains
December 16, 2024 at 1:18 AM
me thinking I covered all my edge cases and forgetting about double top level country appended domains
Great idea I'm never going to get around to actually doing:
Benchmarking models based on non-oracle and expert tests.
I want to see which LLMs can parse documentation, find typos in code, scrub data and adhere to defined output instructions.
Benchmarking models based on non-oracle and expert tests.
I want to see which LLMs can parse documentation, find typos in code, scrub data and adhere to defined output instructions.
I really wish there was more interest in benchmarking AI models on real tasks.
The fact that there are not 30 different benchmarks from different organizations in medicine, in law, in advice quality, etc. is a big shame. People are using systems for these things anyway & we don’t know implications.
The fact that there are not 30 different benchmarks from different organizations in medicine, in law, in advice quality, etc. is a big shame. People are using systems for these things anyway & we don’t know implications.
December 16, 2024 at 12:48 AM
Great idea I'm never going to get around to actually doing:
Benchmarking models based on non-oracle and expert tests.
I want to see which LLMs can parse documentation, find typos in code, scrub data and adhere to defined output instructions.
Benchmarking models based on non-oracle and expert tests.
I want to see which LLMs can parse documentation, find typos in code, scrub data and adhere to defined output instructions.
Thinking about how VB .NET + Visual Studio 2002.
I hated it so much, but it was the foundation for CS degrees back then. It pretty much derailed my enthusiasm for coding for a decade and convinced me for just as long I didn't actually want to code.
I hated it so much, but it was the foundation for CS degrees back then. It pretty much derailed my enthusiasm for coding for a decade and convinced me for just as long I didn't actually want to code.
December 15, 2024 at 11:10 PM
Thinking about how VB .NET + Visual Studio 2002.
I hated it so much, but it was the foundation for CS degrees back then. It pretty much derailed my enthusiasm for coding for a decade and convinced me for just as long I didn't actually want to code.
I hated it so much, but it was the foundation for CS degrees back then. It pretty much derailed my enthusiasm for coding for a decade and convinced me for just as long I didn't actually want to code.
Man, Replit AI just really wants to be a font end designer.
Every prompt I give it, even if it's a backend database tool, it ends up building some kind of React web frontend.
Just gave up and leaned into it. Use it for MVP porting of local scripts into small web apps 🤷
Every prompt I give it, even if it's a backend database tool, it ends up building some kind of React web frontend.
Just gave up and leaned into it. Use it for MVP porting of local scripts into small web apps 🤷
December 15, 2024 at 11:06 PM
Man, Replit AI just really wants to be a font end designer.
Every prompt I give it, even if it's a backend database tool, it ends up building some kind of React web frontend.
Just gave up and leaned into it. Use it for MVP porting of local scripts into small web apps 🤷
Every prompt I give it, even if it's a backend database tool, it ends up building some kind of React web frontend.
Just gave up and leaned into it. Use it for MVP porting of local scripts into small web apps 🤷
Would encourage everyone in AI space to take a look at Bain's 2024 Generative AI Surveys.
Adoption tends and concerns from companies is a revelation.
People are less worried about safety, much more worried that almost all the tools being sold/pitched are low quality.
Adoption tends and concerns from companies is a revelation.
People are less worried about safety, much more worried that almost all the tools being sold/pitched are low quality.
December 15, 2024 at 10:44 PM
Would encourage everyone in AI space to take a look at Bain's 2024 Generative AI Surveys.
Adoption tends and concerns from companies is a revelation.
People are less worried about safety, much more worried that almost all the tools being sold/pitched are low quality.
Adoption tends and concerns from companies is a revelation.
People are less worried about safety, much more worried that almost all the tools being sold/pitched are low quality.
Having a blind spot, where I checked out of learning new dev stuff for about a decade (2011-2018 ish)...
I'm still baffled by how much modern stacks are so damn bloated and web based.
Not bad just a alien strange world where everything is JavaScript
I'm still baffled by how much modern stacks are so damn bloated and web based.
Not bad just a alien strange world where everything is JavaScript
December 15, 2024 at 10:34 PM
Having a blind spot, where I checked out of learning new dev stuff for about a decade (2011-2018 ish)...
I'm still baffled by how much modern stacks are so damn bloated and web based.
Not bad just a alien strange world where everything is JavaScript
I'm still baffled by how much modern stacks are so damn bloated and web based.
Not bad just a alien strange world where everything is JavaScript
Meme that only makes sense for outcomes analysts.
Blind has not helped the problem of people understanding what "total comp" actually means.
Blind has not helped the problem of people understanding what "total comp" actually means.
December 15, 2024 at 7:43 PM
Meme that only makes sense for outcomes analysts.
Blind has not helped the problem of people understanding what "total comp" actually means.
Blind has not helped the problem of people understanding what "total comp" actually means.
Don't have an alt on here for my data science thoughts, so guess I'll be rolling up my semi-anon work and posting into my IRL account.
Likely overdue, but means my saltier takes will need to be reigned in a bit
Likely overdue, but means my saltier takes will need to be reigned in a bit
December 15, 2024 at 7:28 PM
Don't have an alt on here for my data science thoughts, so guess I'll be rolling up my semi-anon work and posting into my IRL account.
Likely overdue, but means my saltier takes will need to be reigned in a bit
Likely overdue, but means my saltier takes will need to be reigned in a bit