Ask me about Reinforcement Learning Research @ Sony AI AI should learn from its experiences, not copy your data. My website for answering RL questions: https://www.decisionsanddragons.com/ Views and posts are my own.https://bsky.app/profile/jmac-ai.bsky.socialhttps://bsky.app/profile/jmac-ai.bsky.social/post/3mllfyt32rs2mThis is neat. Between this and SigReg, are random 1d projections having a moment? [contains quote post or other embedded content]11 May 2026 13:47 +0000at://did:plc:74fhcdv7jj53pu254mlql3kr/app.bsky.feed.post/3mllfyt32rs2mhttps://bsky.app/profile/jmac-ai.bsky.social/post/3mlbdv7sfyk2iPlotting a variable against itself with a weird placement of axes. [contains quote post or other embedded content]07 May 2026 13:43 +0000at://did:plc:74fhcdv7jj53pu254mlql3kr/app.bsky.feed.post/3mlbdv7sfyk2ihttps://bsky.app/profile/jmac-ai.bsky.social/post/3mlbcdg6rnc2tOne of my uses for LLMs is to feed it a paper and have it help untangle abuse of mathematical notation or explain otherwise unfamiliar notation from an adjacent community. It's so much nicer than googling that stuff.07 May 2026 13:15 +0000at://did:plc:74fhcdv7jj53pu254mlql3kr/app.bsky.feed.post/3mlbcdg6rnc2thttps://bsky.app/profile/jmac-ai.bsky.social/post/3mjazcb7v3222Something I find kind of unique to Google's Gemini models is when "discussing" a topic, they will often end their turn by asking what I think about some other relationship. This is actually awesome because one of my main uses of LLMs is to work through ideas/thoughts like a rubber duck++11 Apr 2026 23:43 +0000at://did:plc:74fhcdv7jj53pu254mlql3kr/app.bsky.feed.post/3mjazcb7v3222https://bsky.app/profile/jmac-ai.bsky.social/post/3mjaiik5j4s2sEntangling AI technology with religion is a colossally bad idea.11 Apr 2026 18:42 +0000at://did:plc:74fhcdv7jj53pu254mlql3kr/app.bsky.feed.post/3mjaiik5j4s2shttps://bsky.app/profile/jmac-ai.bsky.social/post/3mja6vn6r2s2wThis is an excellent post. I've made similar arguments against doomerism before, but I never made the connection to it being another flavor of the "counting argument" even though I've argued against its other flavors before too! This framing exposes the flaw very clearly. [contains quote post or other embedded content]11 Apr 2026 15:50 +0000at://did:plc:74fhcdv7jj53pu254mlql3kr/app.bsky.feed.post/3mja6vn6r2s2whttps://bsky.app/profile/jmac-ai.bsky.social/post/3mj656bqwjc2nI’m just going to refer to this next time people tell me LLMs are experiencing human like emotions purely because it appears to. [contains quote post or other embedded content]10 Apr 2026 20:14 +0000at://did:plc:74fhcdv7jj53pu254mlql3kr/app.bsky.feed.post/3mj656bqwjc2nhttps://bsky.app/profile/jmac-ai.bsky.social/post/3miw343hfqk2wTrump starts an illegal war. Then he threatens to destroy an entire civilization. Yet that's still not enough for congress to stop the illegal war. It's still not enough for the press to sound the loudest alarms every chance they get. The moral failing of this country is staggering.07 Apr 2026 15:16 +0000at://did:plc:74fhcdv7jj53pu254mlql3kr/app.bsky.feed.post/3miw343hfqk2whttps://bsky.app/profile/jmac-ai.bsky.social/post/3mivvcbryj22hIt's common in RL papers to compare against 5+ baselines. Is this useful or checkbox ticking? Frequently, one good SOTA baseline seems enough for me. Adding additional baselines only seems useful if they communicate some idea new the other baseline(s) don't.07 Apr 2026 13:32 +0000at://did:plc:74fhcdv7jj53pu254mlql3kr/app.bsky.feed.post/3mivvcbryj22hhttps://bsky.app/profile/jmac-ai.bsky.social/post/3mily3dhbss2wI think artificial perception would have a step function in quality if we can figure out how to train indefinite computation on percepts. That is, the analog to how RL on coding/math/etc. allows indefinite computation on prompts for LLMs. The same goes for everything else.03 Apr 2026 14:55 +0000at://did:plc:74fhcdv7jj53pu254mlql3kr/app.bsky.feed.post/3mily3dhbss2whttps://bsky.app/profile/jmac-ai.bsky.social/post/3mielhwuew22bIt's an underappreciated fact that auto-regressive sequence models over discrete outputs (e.g., an LLM) _are_ hybrid neruo-symbolic systems. And this fact is crucial in making them powerful (unlocks Turing completeness)31 Mar 2026 16:21 +0000at://did:plc:74fhcdv7jj53pu254mlql3kr/app.bsky.feed.post/3mielhwuew22bhttps://bsky.app/profile/jmac-ai.bsky.social/post/3miegn34cic2gThe real bitter lesson is the rest of the AI community slowly conceding how important RL is to the AI problem.31 Mar 2026 14:54 +0000at://did:plc:74fhcdv7jj53pu254mlql3kr/app.bsky.feed.post/3miegn34cic2ghttps://bsky.app/profile/jmac-ai.bsky.social/post/3mibszyonps23Maybe instead of eating up all the memory, we focus that money on research to close the remaining gap between transformers and RNNs? [contains quote post or other embedded content]30 Mar 2026 13:58 +0000at://did:plc:74fhcdv7jj53pu254mlql3kr/app.bsky.feed.post/3mibszyonps23https://bsky.app/profile/jmac-ai.bsky.social/post/3mi7okv3mlc2fGreat talk. A high-level takeway for the "LLMs are useless" crowd is, no they are not, and it's dangerous reality denial to hold that position. Also, love that "rewrite it Rust" made an appearance :p [contains quote post or other embedded content]29 Mar 2026 17:33 +0000at://did:plc:74fhcdv7jj53pu254mlql3kr/app.bsky.feed.post/3mi7okv3mlc2fhttps://bsky.app/profile/jmac-ai.bsky.social/post/3mi2kph6jfc2tUntil now, I don't think I've ever lived through an extended time period where it's better to electronics sooner than later because the prices would only get meaningfully worse. It sucks.27 Mar 2026 16:41 +0000at://did:plc:74fhcdv7jj53pu254mlql3kr/app.bsky.feed.post/3mi2kph6jfc2thttps://bsky.app/profile/jmac-ai.bsky.social/post/3mi2ax4dios2hI consider myself a poor writer, but this is similar to how I use LLMs to improve my writing. Not _do_ my writing. Improve it. I'll find an awkward sentence that I wrote and iterate with an LLM 10 times on it. And then I learn from the experience to avoid the problem in the first place. [contains quote post or other embedded content]27 Mar 2026 13:46 +0000at://did:plc:74fhcdv7jj53pu254mlql3kr/app.bsky.feed.post/3mi2ax4dios2hhttps://bsky.app/profile/jmac-ai.bsky.social/post/3mhvtvylt7k2nWe absolutely need regulation on how AI technology is used/made. "Pause developing AI models 'more powerful than' GPT4" is an abysmal poorly-formulated policy that if anything entrenches the power of the current abusive leaders. How about you instead work on useful policies to stop bad behavior? [contains quote post or other embedded content]25 Mar 2026 19:42 +0000at://did:plc:74fhcdv7jj53pu254mlql3kr/app.bsky.feed.post/3mhvtvylt7k2n