Alex J. Champandard 🍂
@alexjc
Followers
26K
Following
17K
Media
4K
Statuses
49K
Building tools and teams where humans ≫ machines. AI, ML, research & development. co-Founded #CreativeAI #⚘
Europe
Joined September 2007
🚀 New Record On #NanoGPT Speed-Run! TokenMonster Edition 🔥 3.28 FineWeb val_loss in 2.77 min, prior record 2.997 minutes! (Detailed explanations in thread.) Special thanks to @LambdaAPI for supporting the research with compute credits. Change Log: - Tokenized FineWeb with
5
3
59
💰Countering Dems hollow words on “affordability” Polls and elections highlight the Republican need to take on Democrats affordability message, and to take their affordability wins from the Big Beautiful Bill on the road.
9
74
308
1/5 Excited to announce our paper on confessions! We train models to honestly report whether they “hacked”, “cut corners”, “sandbagged” or otherwise deviated from the letter or spirit of their instructions. @ManasJoglekar Jeremy Chen @GabrielDWu1 @jasonyo @j_asminewang
7
20
153
Today in Current Affairs, professor Ron Purser exposes how AI's destruction of the university is even worse than you think, and goes well beyond students cheating with ChatGPT: https://t.co/TAUD4R7gxK
223
2K
8K
An 'up-to-the-minute' portrait of our universal electromagnetic superpositional cosmos of creation. I was fortunate to be one of the guinea pigs for the entire 1969-1973 'mod' school experiment at Nordhoff High School, Ojai CA, where the students planned and scheduled their own
5
8
49
Fukushima's video (1986) shows a CNN that recognises handwritten digits [3], three years before LeCun's video (1989). CNN timeline taken from [5]: ★ 1969: Kunihiko Fukushima published rectified linear units or ReLUs [1] which are now extensively used in CNNs. ★ 1979:
Yann LeCun’s 1989 convolutional neural network demo, the foundation for the CNNs we still use today. It's amazing how far we've come since then!
68
483
3K
Early stages of a project from a year ago that never came to be.
6
7
104
If the initial benchmarks scores (and graphs used for PR) showcased a 3x reduction in size for the same performance, I think the broader public reception would have been less tepid. Just looking at this, it just seems 7% behind other existing models in the 4.x series...
Small PMPP-Eval update for freshly released Intellect-3 by @PrimeIntellect From my personal tests it was clear that its outperforming the Air variant (which uses same base model) numbers are confirming this with +%34 difference compared to Air and on par with 3x sized GLM-4.5
0
0
1
f=0,draw=r=>{for(f||createCanvas(W=400,W),noStroke(background(0)),o=7.5,t=300,y=-t;y<t;y+=35){for(Y=y+f%70,x=-t;x<t;x+=35)for(i=0;i<35;i+=1)v=Y-sin(m=abs(i*(150-mag(x,Y)))/99)*m/2,h=x+cos(m)*m/2,fill(9*m),circle(h+200+o,v+210,3);o=-o}f+=.3};//#つぶやきProcessing #p5js
4
38
227
?
ChatGPT Android app 1.2025.329 beta includes new references to an "ads feature" with "bazaar content", "search ad" and "search ads carousel"
330
2K
55K
One of my studies on electromagnetic sculpting from a couple of years ago. Exploring magnetic fields and the emergent interaction between matter and electricity.
60
244
2K
opus 4.5 marked the jump from sde1 to sde2. i’m having a lot of fun and getting a lot of shit done but i’m also realizing we’re going to be the last generation of software engineers
101
66
2K
Robotaxis will be in the highest demand. No one will own their own vehicle. Also robotaxis:
94
41
794
A recent client win: 10X ROAS on whitelisting ads INSANE
6
1
36
y'all in the 'jagged frontier' debate are missing something:
79
176
4K
A latent process that operates over plans? I've been working on this recently! What's most fascinating to me: my 6k parameter system can match aspects of models 100,000x bigger. Scaling laws apply very differently too... 🤔
11/ People often say “humans think like diffusion, not left-to-right.” We agree humans aren’t strict L2R generators, but current diffusion LLMs still miss a core part of human cognition. Humans don’t reason in absolute token slots (“what’s the 25th word from now?”). We think
0
0
4
recoiling so far away from the act of creation that you accidentally invent the radio station
39
1K
20K
I looked into this and the answer is so funny. In the No Thinking setting, Opus 4.5 repurposes the Python tool to have an extended chain of thought. It just writes long comments, prints something simple, and loops! Here's how it starts one problem:
Opus 4.5 scores the same on FrontierMath regardless of thinking budget, in contrast to GPT-5.1 where higher reasoning settings correspond to higher scores. However, on OTIS Mock AIME, another math benchmark, we see the thinking budget make a difference for Opus 4.5 as well.
23
49
764
Use only one animation to convince people to follow you
165
4K
43K
For Gemini 3, I don't rule out bugs in @cursor_ai — as many new features don't work, worktrees getting trashed or even renamed (!) mid-way through agent working. But since Opus 4.5 manages around those bugs, it can't be entirely on the Cursor side.
0
0
1
My verdict is that it's significantly better than Gemini 3. It's at least as smart and just got more polish to it. Alignment on little details also significantly higher. Gemini 3 gets many things mixed up after a half-dozen messages, and completely confused after compaction.
1
0
1