Lon
@Lon
Followers
3K
Following
2K
Media
12
Statuses
362
Absurdist Intern in Satirical Studies. *tweets randomly self-destruct
Joined February 2007
I'm still arguing with ChatGPT5 just as often as ChatGPT4 over stupid shit as it continues to permute and twist around facts and details until landing in surreal lala land where it confidently asserts up is down and down is up.
0
0
1
The danger in LLM overconfidence is it heavily biases from starting point of conversation context. For ex, with medical and mental health topics; given the same information LLM's will give you an overconfidently different result based on convo order alone.
First o3 pro usage. Many mistakes. Massive overconfidence. Clear inability to distinguish citations, pay attention to dates. Does anyone else actually use these models? They may be smarter on paper but they are increasingly lazy and evil in practice.
0
0
1
I gave this image to o3 geoguessr and it was able to find the location within 6 feet.
0
0
2
When we started, .we thought we were .out to build .an intelligence system .and convince the world .that we'd done it. We're now starting .to wonder whether .we're gonna build systems .that we're not convinced .are fully intelligent, .and we're trying to convince .the world that.
0
0
1
*taps sign*
From the point of view of science I've never been too excited about the Game of Life. But now I realize that what's really exciting about it is what it tells us about what we can call metaengineering. That half century of impressive Life hacking gives us a uniquely clean example
0
0
5
Just working off my debt at the Loyalty Center. .
Something about this imagery is deeply unsettling, but it's actually really interesting. Quest 2s have entered select California prisons for inmates sent to solitary confinement. Crazy part is that after VR "programming", inmate infractions fell 96%.
0
0
1
Not cool how they are doing my boy, Ben.
Totally false. Untrue. These are lies coming from Russian media. I completely self-funded my humanitarian trip to Ukraine. There was no funding from USAID and certainly no personal payments.
0
0
2
We know from every LLM trained thus far that given a long enough generated sequence length every auto-regressive transformer decoheres into garbage. That aligns well with Yannโ assertion. All of the RL / internal CoT monologue is to keep it within the error boundary for longer.
0
0
0
I would like someone to explain which facet of MCTS is missing in o3โs capabilities, as if recursive CoT isnโt equivalent to rollout and there isnโt already an implicit policy function and backtracking in play in the weights.
1
0
0
People are dunking on Yann Lecun while claiming that o3 is leading to AGI using only auto-regressive LLMs, but leaving out the part where MCTS is being implicitly trained into the weights through RL and exponential regressions are being used to generate correct responses.
1
0
4
The juxtaposition of the humans-as-cattle Beast Games and the announcement of OpenAI o3 is really hitting all of the Black Mirror dystopian full-send vibes for our future.
0
1
3
I'm just an intern, but I think this means that charges should now be dropped against Roman Storm and Roman Semenov.
Privacy wins. Today the Fifth Circuit held that @USTreasuryโs sanctions against Tornado Cash smart contracts are unlawful. This is a historic win for crypto and all who cares about defending liberty. @coinbase is proud to have helped lead this important challenge. 1/6.
1
0
6
