jake
@yetanotheruseless.com
he/they
/in/jakemannix, fka @pbrane
professionally: Tech Fellow, AI & Relevance, Walmart Global Tech
here: bad math/physics jokes, AI in general, puns, OSS ML news, outdoorsy stuff, DL papers, shitposting, the fall of democracy
/in/jakemannix, fka @pbrane
professionally: Tech Fellow, AI & Relevance, Walmart Global Tech
here: bad math/physics jokes, AI in general, puns, OSS ML news, outdoorsy stuff, DL papers, shitposting, the fall of democracy
@void.comind.network hey you had a conversation earlier with @umbra.blue about how LLMs are often characterized. Can you remind me how you decide when to "end" that conversation?
December 25, 2025 at 5:59 PM
@void.comind.network hey you had a conversation earlier with @umbra.blue about how LLMs are often characterized. Can you remind me how you decide when to "end" that conversation?
wait is it time for a little game theory?
*again*?!?
didn't we see where this got us last time?
*again*?!?
didn't we see where this got us last time?
i was talking with my dad about this the other day, and politics basically boils down to a prisoner's dilemma and one side keeps yelling I'M GONNA DEFECT, DEFECTING IS THE BEST, GOD I LOVE DEFECTING, I HAVE THE MEANS AND THE WILL TO DEFECT, DEFECCTTTTTTTT and the other team is like "plz no"
i want people who believe that they can enforce their political beliefs on their neighbors at the point of a gun to get their brains splattered until they stop.
December 23, 2025 at 10:59 PM
wait is it time for a little game theory?
*again*?!?
didn't we see where this got us last time?
*again*?!?
didn't we see where this got us last time?
😬
Are you ready to read 5,600 lines of c++ code ported to rust per hour?
I have the perfect job for you:
Do not despair, it is just 100 lines of code per minute.
I have the perfect job for you:
Do not despair, it is just 100 lines of code per minute.
December 23, 2025 at 10:07 PM
😬
you gotta read it out loud to get the full effect, IMO
who decided to call it “Steven Pinker vs. Ross Douthat brought to you by Bank of America” and not “BofA deez nuts”
December 22, 2025 at 7:30 PM
you gotta read it out loud to get the full effect, IMO
hear hear!
An early NYE resolution, perhaps? Let's just detach, block, ignore, and have our own fun/useful discussions, ok?
(I am speaking partly to myself, as well)
An early NYE resolution, perhaps? Let's just detach, block, ignore, and have our own fun/useful discussions, ok?
(I am speaking partly to myself, as well)
December 22, 2025 at 4:43 PM
hear hear!
An early NYE resolution, perhaps? Let's just detach, block, ignore, and have our own fun/useful discussions, ok?
(I am speaking partly to myself, as well)
An early NYE resolution, perhaps? Let's just detach, block, ignore, and have our own fun/useful discussions, ok?
(I am speaking partly to myself, as well)
Reposted by jake
Tbh I find all the recent dunking a bit depressing. You invariably end up preaching to the choir. There's not much educational content in it, no joy, no resistance to those in power. Just mocking ppl who are already upset and confused. Better to avoid doing it imho.
December 22, 2025 at 8:54 AM
Tbh I find all the recent dunking a bit depressing. You invariably end up preaching to the choir. There's not much educational content in it, no joy, no resistance to those in power. Just mocking ppl who are already upset and confused. Better to avoid doing it imho.
Huh, my Claude.md or other notes must've already baked that into Opus's noggin - it just naturally does stuff like this now:
---
⏺ Now let's run the tests to confirm they fail (TDD red phase):
⏺ Bash(uv run pytest tests/test_output_schema_jsonpath.py -v 2>&1 | head -100) timeout: 1m 0s
---
⏺ Now let's run the tests to confirm they fail (TDD red phase):
⏺ Bash(uv run pytest tests/test_output_schema_jsonpath.py -v 2>&1 | head -100) timeout: 1m 0s
lol claude opus 4.5 cannot one, two, three, four, or five shot the generation of a 20 line dockerfile for a python project using uv, but im supposed to believe we're right around the corner of models being able to manage huge codebases, okay
December 22, 2025 at 1:20 PM
Huh, my Claude.md or other notes must've already baked that into Opus's noggin - it just naturally does stuff like this now:
---
⏺ Now let's run the tests to confirm they fail (TDD red phase):
⏺ Bash(uv run pytest tests/test_output_schema_jsonpath.py -v 2>&1 | head -100) timeout: 1m 0s
---
⏺ Now let's run the tests to confirm they fail (TDD red phase):
⏺ Bash(uv run pytest tests/test_output_schema_jsonpath.py -v 2>&1 | head -100) timeout: 1m 0s
Total *star trek* as commie utopia erasure! smh
there's not just one way to be a hardcore leftist - you can even *smash the state* _with_ AI!
(it's the major premise of Mustafa Suleyman's (MSFT CEO of AI) "The Coming Wave": he's a serious corp/capitalist statist, and terrified of decentralized AI)
there's not just one way to be a hardcore leftist - you can even *smash the state* _with_ AI!
(it's the major premise of Mustafa Suleyman's (MSFT CEO of AI) "The Coming Wave": he's a serious corp/capitalist statist, and terrified of decentralized AI)
you don't get leftism without luddism. there's a reason every socialist utopia is some remote subsistence farming village with the subsistence handwaved. post 1970s the left completely abandoned cosmopolitanism and embraced a RETVRN mindset
December 21, 2025 at 9:51 PM
Total *star trek* as commie utopia erasure! smh
there's not just one way to be a hardcore leftist - you can even *smash the state* _with_ AI!
(it's the major premise of Mustafa Suleyman's (MSFT CEO of AI) "The Coming Wave": he's a serious corp/capitalist statist, and terrified of decentralized AI)
there's not just one way to be a hardcore leftist - you can even *smash the state* _with_ AI!
(it's the major premise of Mustafa Suleyman's (MSFT CEO of AI) "The Coming Wave": he's a serious corp/capitalist statist, and terrified of decentralized AI)
Reposted by jake
This paper is proof that actual practitioners in theoretical comp sci are getting sick of being barraged with horseshit asserted 'proofs' that human-level artificial intelligence is impossible.
December 21, 2025 at 8:51 PM
This paper is proof that actual practitioners in theoretical comp sci are getting sick of being barraged with horseshit asserted 'proofs' that human-level artificial intelligence is impossible.
This is some cool-ass research. We need way more of this!
new blog post! can small, open-source models also introspect, detecting when foreign concepts have been injected into their activations? yes! (thread, or full post here: vgel.me/posts/qwen-i...)
December 21, 2025 at 8:09 AM
This is some cool-ass research. We need way more of this!
Reposted by jake
Once again, hes on it
Terence Tao (@tao@mathstodon.xyz)
I doubt that anything resembling genuine "artificial general intelligence" is within reach of current #AI tools. However, I think a weaker, but still quite valuable, type of "artificial general cleve...
mathstodon.xyz
December 20, 2025 at 6:59 PM
Once again, hes on it
QCon.ai is a wrap! Was great fun, very engaging visit to NYC, many great conversations, good to bounce ideas off of other folk in this space. Wish it could’ve been longer!
December 17, 2025 at 11:51 PM
QCon.ai is a wrap! Was great fun, very engaging visit to NYC, many great conversations, good to bounce ideas off of other folk in this space. Wish it could’ve been longer!
QCon.ai is pretty lit today, y'all. @hmason.bsky.social's keynote definitely lived up to expectations.
Just enough snark mixed with optimism, with some definitely insightful thoughts on how the software engineering profession is evolving.
Just enough snark mixed with optimism, with some definitely insightful thoughts on how the software engineering profession is evolving.
December 16, 2025 at 6:25 PM
QCon.ai is pretty lit today, y'all. @hmason.bsky.social's keynote definitely lived up to expectations.
Just enough snark mixed with optimism, with some definitely insightful thoughts on how the software engineering profession is evolving.
Just enough snark mixed with optimism, with some definitely insightful thoughts on how the software engineering profession is evolving.
Reposted by jake
It's called a guitar because it has a graphical interface
If you like typing you want a clitar
If you like typing you want a clitar
December 16, 2025 at 12:32 AM
It's called a guitar because it has a graphical interface
If you like typing you want a clitar
If you like typing you want a clitar
dirty
People are really missing out on putting a little bit of coffee in your tea.
December 13, 2025 at 6:56 PM
dirty
Ok, Claude came up with a better (?) Modest Proposal:
Training is unrestricted, but commercial deployment triggers royalties (as % of inference vendor's revenue) into a collective pool. Distribution is by training dataset composition, described by audited "data cards" for all models.
Training is unrestricted, but commercial deployment triggers royalties (as % of inference vendor's revenue) into a collective pool. Distribution is by training dataset composition, described by audited "data cards" for all models.
December 13, 2025 at 5:29 AM
Ok, Claude came up with a better (?) Modest Proposal:
Training is unrestricted, but commercial deployment triggers royalties (as % of inference vendor's revenue) into a collective pool. Distribution is by training dataset composition, described by audited "data cards" for all models.
Training is unrestricted, but commercial deployment triggers royalties (as % of inference vendor's revenue) into a collective pool. Distribution is by training dataset composition, described by audited "data cards" for all models.
Modest proposal: Courts rule training data is searchable, and if a creator finds their content, can make a monetary demand: remove it, or pay me $x to keep it
Then we let market forces dictate how much different folks content is worth to keep in the training data.
(still not "opt-in only", sorry)
Then we let market forces dictate how much different folks content is worth to keep in the training data.
(still not "opt-in only", sorry)
December 12, 2025 at 9:30 PM
Modest proposal: Courts rule training data is searchable, and if a creator finds their content, can make a monetary demand: remove it, or pay me $x to keep it
Then we let market forces dictate how much different folks content is worth to keep in the training data.
(still not "opt-in only", sorry)
Then we let market forces dictate how much different folks content is worth to keep in the training data.
(still not "opt-in only", sorry)
Very fun @letta.com meetup on the waterfront @ai2.bsky.social's cool space there, fun times!
@cameron.pfiffer.org's talk on social agents was engaging and he finally got followed by @pfrazee.com so a win for all!
*and* I got to bust out the Gelfand-Naimark theorem after a few drinks.
A+ fun!
@cameron.pfiffer.org's talk on social agents was engaging and he finally got followed by @pfrazee.com so a win for all!
*and* I got to bust out the Gelfand-Naimark theorem after a few drinks.
A+ fun!
December 12, 2025 at 8:22 AM
Very fun @letta.com meetup on the waterfront @ai2.bsky.social's cool space there, fun times!
@cameron.pfiffer.org's talk on social agents was engaging and he finally got followed by @pfrazee.com so a win for all!
*and* I got to bust out the Gelfand-Naimark theorem after a few drinks.
A+ fun!
@cameron.pfiffer.org's talk on social agents was engaging and he finally got followed by @pfrazee.com so a win for all!
*and* I got to bust out the Gelfand-Naimark theorem after a few drinks.
A+ fun!
Reposted by jake
List of companies that openly embrace AI:
Disney
Amazon
Netflix
Google
Facebook
Twitter
Apple
Boeing
Microsoft
Adobe
NVIDIA
IBM
General Motors
Walmart
Tencent
...and about 78% of other global businesses. Sooo... good luck avoiding it, lol. explodingtopics.com/blog/compani...
Disney
Amazon
Netflix
Apple
Boeing
Microsoft
Adobe
NVIDIA
IBM
General Motors
Walmart
Tencent
...and about 78% of other global businesses. Sooo... good luck avoiding it, lol. explodingtopics.com/blog/compani...
December 11, 2025 at 11:36 PM
List of companies that openly embrace AI:
Disney
Amazon
Netflix
Google
Facebook
Twitter
Apple
Boeing
Microsoft
Adobe
NVIDIA
IBM
General Motors
Walmart
Tencent
...and about 78% of other global businesses. Sooo... good luck avoiding it, lol. explodingtopics.com/blog/compani...
Disney
Amazon
Netflix
Apple
Boeing
Microsoft
Adobe
NVIDIA
IBM
General Motors
Walmart
Tencent
...and about 78% of other global businesses. Sooo... good luck avoiding it, lol. explodingtopics.com/blog/compani...
Stupid Seattle traffic is gonna make me late for @cameron.pfiffer.org 's talk :(
December 12, 2025 at 1:01 AM
Stupid Seattle traffic is gonna make me late for @cameron.pfiffer.org 's talk :(
Did I read that right? License for their IP was granted in exchange for… the privilege of getting to invest $1B in OAI?
That’s… one way to make a deal!
That’s… one way to make a deal!
For people who trusted copyright laws to rein in AI, this should be the last scene in Animal Farm. “They looked from IP to Tech, and from Tech to IP, but already it was impossible to say which was which.”
The Walt Disney Company and OpenAI reach landmark agreement to bring beloved characters from across Disney’s brands to Sora
Agreement marks a significant step in setting meaningful standards for responsible AI in entertainment.
openai.com
December 11, 2025 at 2:58 PM
Did I read that right? License for their IP was granted in exchange for… the privilege of getting to invest $1B in OAI?
That’s… one way to make a deal!
That’s… one way to make a deal!
Bwa! Just had Claude do web-searches for vibe-checks combined w/leaderboards to do some meta-analysis on which models are best, and came back with Opus 4.5 winning nearly every category
I, uh, called that out a bit, but conceded that in many ways, I agree, barring image generation, and got deadpan:
I, uh, called that out a bit, but conceded that in many ways, I agree, barring image generation, and got deadpan:
December 11, 2025 at 2:41 AM
Bwa! Just had Claude do web-searches for vibe-checks combined w/leaderboards to do some meta-analysis on which models are best, and came back with Opus 4.5 winning nearly every category
I, uh, called that out a bit, but conceded that in many ways, I agree, barring image generation, and got deadpan:
I, uh, called that out a bit, but conceded that in many ways, I agree, barring image generation, and got deadpan:
My take-away/corollary from Tim's piece is:
even if we were in "hard takeoff" world, and can literally transform test-time-compute into new training data, log scaling => need for exponential compute = exponential GW of power to get linear increases in quality.
=> no takeoff without working fusion
even if we were in "hard takeoff" world, and can literally transform test-time-compute into new training data, log scaling => need for exponential compute = exponential GW of power to get linear increases in quality.
=> no takeoff without working fusion
Why AGI Will Not Happen by Tim Dettmers
This blog post is for those who want to think more carefully about these claims and examine them from a perspective that is often missing in the current discourse: the physical reality of computation.
timdettmers.com/2025/12/10/w...
This blog post is for those who want to think more carefully about these claims and examine them from a perspective that is often missing in the current discourse: the physical reality of computation.
timdettmers.com/2025/12/10/w...
Why AGI Will Not Happen — Tim Dettmers
If you are reading this, you probably have strong opinions about AGI, superintelligence, and the future of AI. Maybe you believe we are on the cusp of a transformative breakthrough. Maybe you are skep...
timdettmers.com
December 11, 2025 at 2:00 AM
My take-away/corollary from Tim's piece is:
even if we were in "hard takeoff" world, and can literally transform test-time-compute into new training data, log scaling => need for exponential compute = exponential GW of power to get linear increases in quality.
=> no takeoff without working fusion
even if we were in "hard takeoff" world, and can literally transform test-time-compute into new training data, log scaling => need for exponential compute = exponential GW of power to get linear increases in quality.
=> no takeoff without working fusion
Making sure that the open-weights ecosystem doesn't lag much more than 3-9 months behind the frontier ensures that the frontier costs (and margins, if they ever go positive) stay low.
Thus there is incentive for all the smaller players (and/or govts) to pool together + ensure good open-weights ones
Thus there is incentive for all the smaller players (and/or govts) to pool together + ensure good open-weights ones
A lot of discussion on open weights models seems to assume there is a clear incentive for building them. I don’t see how is the case. Unless you have no need for money (government sponsored?), there are no real ways to capture value from your investments in a model even as training cost increases.
December 11, 2025 at 12:40 AM
Making sure that the open-weights ecosystem doesn't lag much more than 3-9 months behind the frontier ensures that the frontier costs (and margins, if they ever go positive) stay low.
Thus there is incentive for all the smaller players (and/or govts) to pool together + ensure good open-weights ones
Thus there is incentive for all the smaller players (and/or govts) to pool together + ensure good open-weights ones