Post
25
CHISEL: Chunk Holdout Influence Scoring by Elimination Loss
Removing pieces to reveal what matters
Link: https://dejan.ai/media/pdf/CHISEL.pdf
Removing pieces to reveal what matters
Link: https://dejan.ai/media/pdf/CHISEL.pdf
Dan Petrovic PRO
dejanseo
AI & ML interests
DEJAN - The AI SEO agency.
[DEJAN](https://dejan.ai/)
Recent Activity
View all activity
Organizations
posted an update about 4 hours ago
posted an update 30 days ago
Post
133
10 Search Subqueries in 200 Microseconds: 1-Bit Consistency Fanout
We built Fanout Diffusion, an ultra-low-latency model that expands a search query into 10 diverse subquery vectors in a single forward pass.
Inspired by the continuous retrieval framework in R4T (arXiv:2603.06397), our goal was to make query expansion fast enough for production search without running language models.
How it works:
• Direct Embedding Space: Takes a query embedded via google/embeddinggemma-300m (768-dim) and predicts 10 distinct subquery vectors simultaneously.
• 1-Step Consistency Denoiser: Generates all 10 slots analytically in a single pass with zero ODE integration loops.
• 1-Bit Hardware Tensor Cores: Quantized ternary weights (-1, 0, +1) running on Ampere/Ada sub-byte PTX instructions with INT4 outer projections.
Key numbers on an RTX 4090:
• Latency: 0.200 ms (ONNX) / 0.329 ms (native C++ CUDA graph)
• Throughput: 417,035 queries/second in batch mode
• Model Size: 1.64 MB (PyTorch QAT) / 32.1 MB (ONNX graph)
• Quality: 0.683 prompt alignment across 540k search queries
Interactive Space:
dejanseo/fanout-diffusion
Model weights, ONNX graph, and C++ engine:
dejanseo/fanout-diffusion
We built Fanout Diffusion, an ultra-low-latency model that expands a search query into 10 diverse subquery vectors in a single forward pass.
Inspired by the continuous retrieval framework in R4T (arXiv:2603.06397), our goal was to make query expansion fast enough for production search without running language models.
How it works:
• Direct Embedding Space: Takes a query embedded via google/embeddinggemma-300m (768-dim) and predicts 10 distinct subquery vectors simultaneously.
• 1-Step Consistency Denoiser: Generates all 10 slots analytically in a single pass with zero ODE integration loops.
• 1-Bit Hardware Tensor Cores: Quantized ternary weights (-1, 0, +1) running on Ampere/Ada sub-byte PTX instructions with INT4 outer projections.
Key numbers on an RTX 4090:
• Latency: 0.200 ms (ONNX) / 0.329 ms (native C++ CUDA graph)
• Throughput: 417,035 queries/second in batch mode
• Model Size: 1.64 MB (PyTorch QAT) / 32.1 MB (ONNX graph)
• Quality: 0.683 prompt alignment across 540k search queries
Interactive Space:
dejanseo/fanout-diffusion
Model weights, ONNX graph, and C++ engine:
dejanseo/fanout-diffusion
- 1 reply
replied to their post about 2 months ago
I would be delighted if that was the case.
replied to their post about 2 months ago
replied to their post about 2 months ago
Correct. 5.3 was my reference text that got matched as the closest. I think they're about to release GLM with a visual tower.
posted an update about 2 months ago
Post
3,055
Ox Alpha is GLM
https://dejan.ai/blog/ox-alpha/
A parameter-free k-nearest-neighbour classifier over Normalized Compression Distance (Lee et al., MobiSys ’24, Eq. 1, built on Jiang et al.'s gzip-based text classifier). NCD compares two texts by how well they compress together. C(s) is the gzip-compressed length of s. Text sharing an author's patterns compresses better together than text from a different author, so the method needs no model weights and no embeddings.
The reference corpus covers 60 prompts (essays, code, emails, dialogue, poetry) answered by five known models: GPT-5.5, Claude Opus 5, Gemini 3.7 Flash, Gemini 3.1 Pro Preview, and GLM-5.3, for 293 reference texts. ox-alpha answered the first 13 of those prompts, plus one additional novel prompt never given to the reference models beforehand, for 14 queries in total. Each query was classified against the reference corpus independently, with a k-nearest-neighbour vote (k=5):
Model ox-alpha samples matched
GLM-5.3 7 / 14
Claude Opus 5 3 / 14
Gemini 3.7 Flash 2 / 14
GPT-5.5 1 / 14
Gemini 3.1 Pro Preview 1 / 14
GLM-5.3 wins at every k tested: 7/14 at k=3, 7/14 at k=5, 6/14 at k=7, 7/14 at k=9. Claude Opus 5 is the consistent second place.
https://dejan.ai/blog/ox-alpha/
A parameter-free k-nearest-neighbour classifier over Normalized Compression Distance (Lee et al., MobiSys ’24, Eq. 1, built on Jiang et al.'s gzip-based text classifier). NCD compares two texts by how well they compress together. C(s) is the gzip-compressed length of s. Text sharing an author's patterns compresses better together than text from a different author, so the method needs no model weights and no embeddings.
The reference corpus covers 60 prompts (essays, code, emails, dialogue, poetry) answered by five known models: GPT-5.5, Claude Opus 5, Gemini 3.7 Flash, Gemini 3.1 Pro Preview, and GLM-5.3, for 293 reference texts. ox-alpha answered the first 13 of those prompts, plus one additional novel prompt never given to the reference models beforehand, for 14 queries in total. Each query was classified against the reference corpus independently, with a k-nearest-neighbour vote (k=5):
Model ox-alpha samples matched
GLM-5.3 7 / 14
Claude Opus 5 3 / 14
Gemini 3.7 Flash 2 / 14
GPT-5.5 1 / 14
Gemini 3.1 Pro Preview 1 / 14
GLM-5.3 wins at every k tested: 7/14 at k=3, 7/14 at k=5, 6/14 at k=7, 7/14 at k=9. Claude Opus 5 is the consistent second place.
- 6 replies
posted an update 3 months ago
Post
150
Early-checkpoint diffusion output retains enough corpus fingerprint for source attribution, before the model can form a coherent sentence.
I'm currently pretraining a masked diffusion language model and the following is output from one of its early checkpoints.
---
competitors of Paramount. Combootsseys announced on September 21 Hden 21. All was on Emb Street Journal. We’re not already full of news given how Cinema says tech moves. “Kitt showed that there are millions of investigating unprecedented positive issues for around us, shifting- on wayside during its “Our CES tour.” The subculture industry has allows hatenders against 47 respects since June, including the animated animated movies season 3, and “Ult with Securecom i Got USA Supreme.” On Friday, the Suicide ruled to stop providing the clifihotic, stupid clam of camps and candles, but there was no nihoke sent to me. Emeruination issued that was not enough to provide a generic explanation of the backlash for two months. But the President was lethally following Measure 2016 years of experimental crowdfunding it never received. The kind of thing that’s held during DJ Day’s Fan Bwn and Suicide is the opposite. It’s a little drama and even thwashing the $4ennig proceeds (he does lame forMiggle is patrol to besting) that often dominate nixing high party campaigns into imagardous, mediocre artistic
I asked Gemini, GPT and Claude to guess the site and they all said the same thing: The Verge.
An LLM would probably do the same with a bag of words by looking at clues in keywords. But it could also say Engadget, TechCrunch, Ars or Gizmodo. And it this case they all locked on a very specific site.
Very interesting.
I'm currently pretraining a masked diffusion language model and the following is output from one of its early checkpoints.
---
competitors of Paramount. Combootsseys announced on September 21 Hden 21. All was on Emb Street Journal. We’re not already full of news given how Cinema says tech moves. “Kitt showed that there are millions of investigating unprecedented positive issues for around us, shifting- on wayside during its “Our CES tour.” The subculture industry has allows hatenders against 47 respects since June, including the animated animated movies season 3, and “Ult with Securecom i Got USA Supreme.” On Friday, the Suicide ruled to stop providing the clifihotic, stupid clam of camps and candles, but there was no nihoke sent to me. Emeruination issued that was not enough to provide a generic explanation of the backlash for two months. But the President was lethally following Measure 2016 years of experimental crowdfunding it never received. The kind of thing that’s held during DJ Day’s Fan Bwn and Suicide is the opposite. It’s a little drama and even thwashing the $4ennig proceeds (he does lame forMiggle is patrol to besting) that often dominate nixing high party campaigns into imagardous, mediocre artistic
I asked Gemini, GPT and Claude to guess the site and they all said the same thing: The Verge.
An LLM would probably do the same with a bag of words by looking at clues in keywords. But it could also say Engadget, TechCrunch, Ars or Gizmodo. And it this case they all locked on a very specific site.
Very interesting.
posted an update 3 months ago
Post
264
Can humans recognize AI generated text?
Test: https://dejan.ai/ai-vs-human/
Results: https://dejan.ai/blog/ai-vs-human-test-results/
Test: https://dejan.ai/ai-vs-human/
Results: https://dejan.ai/blog/ai-vs-human-test-results/
- 2 replies
posted an update 10 months ago
Post
528
Loss Curve Forecaster — Predict Where Your Training Is Heading
dejanseo/loss-target-estimator
Ever stared at a loss curve wondering:
- How many more hours until I hit my target loss?
- What will my loss be if I train for 10 more hours?
- Is this thing ever going to converge?
I built a tool to answer these questions.
Upload a WandB CSV export and get instant predictions using power law curve fitting — the same dynamics that govern LLM training scaling laws.
What you get:
Asymptote prediction (your loss floor)
Time-to-target estimates
Progress percentage toward convergence
Dual-axis chart showing both wall time and steps
Tip: Export from WandB with Wall Time as the X-axis to see both time and steps on the chart.
https://dejan.ai/tools/loss-estimator/
dejanseo/loss-target-estimator
Ever stared at a loss curve wondering:
- How many more hours until I hit my target loss?
- What will my loss be if I train for 10 more hours?
- Is this thing ever going to converge?
I built a tool to answer these questions.
Upload a WandB CSV export and get instant predictions using power law curve fitting — the same dynamics that govern LLM training scaling laws.
What you get:
Asymptote prediction (your loss floor)
Time-to-target estimates
Progress percentage toward convergence
Dual-axis chart showing both wall time and steps
Tip: Export from WandB with Wall Time as the X-axis to see both time and steps on the chart.
https://dejan.ai/tools/loss-estimator/