David (@dleeships) profile picture
Bay Area, CA
Joined September 2025
Photo by David (@dleeships)Photo by David (@dleeships)Photo by David (@dleeships)Photo by David (@dleeships)Photo by David (@dleeships)Photo by David (@dleeships)Photo by David (@dleeships)Photo by David (@dleeships)Photo by David (@dleeships)Photo by David (@dleeships)
Critique of @vercel token pass: you need to be on the Pro plan ($ 20) to buy the $30 token pass add-on So it's really paying $50 for $150 (prev was $80)
1
2
456
One thing I’ve been thinking about after fine-tuning Qwen on my cold emails is sample efficiency With only 8–15 training examples, the way I wrote each example mattered a TON. Adding more emails DIDN'T consistently improve the model...
1
1
125
On 30 fresh briefs, the new tune kept the question in every draft. An LLM judge ranked it first against the earlier tune and base Qwen in 25/30 comparisons!
1
15
This was a small SFT experiment, but it gave me a more practical appreciation for sample efficiency. I’d spend more time checking what my existing examples demonstrate before collecting more I’m curious how much further a small, carefully constructed dataset could take this
4
Introducing Speculative Self-Distillation (SSD, NeurIPS 2026), which pushes the accuracy–efficiency–forgetting frontier of self-distillation. In SSD, student and teacher (the same model with additional information) write the rollouts together. Like a private tutor, the teacher lets the student write and takes over only when the student drifts too far. As a result, SSD trains on rollouts that stay as close to the student's own writing (on-policy) as possible while the teacher keeps them on track. 📉 38–69% fewer training tokens than on-policy, matching its accuracy ⏱️ Reaches on-policy accuracy in ~130 vs ~414 min (Science Q&A) 🧠 Forgets less in a stress test: −1.4 vs −4.5 points for on-policy
33
64
18
509
59,395
Post-training Qwen to write cold emails was surprisingly... bad Having it learn about me, about the recipient, and how to read the right context to build was very difficult
1
1
132
Example: I gave it an email I wrote that offered someone at OpenAI my AI memory research, tied to a pricing idea A test example to a recipient from Vercel, but the tuned model reused the pitch + implied I'd worked at Vercel at that time
1
39
Adding context exposed another problem: My context brief included summaries of other people's research, and the model started confusing that with my experience Thus, I labeled: my experience recipient's work third-party research Better context eng improved the results A LOT
16
David retweeted
Big news: Claude Fable 5.2 (Max) by @AnthropicAI takes #1 in Agent Arena with an unprecedented lead: +30.4% net improvement, about 1.6x the #2 model Opus-5.5, across 4.7k real-world agentic sessions from our global community. It's also #1 in Code, Chat and Work. A major leap over Claude Fable 5.1 (Max) - Rank: #5 → #1, Net improvement: +14.8% → +30.4% - Median cost per task: $7.7 → $7.3 (6% lower). The API pricing $10/$50 → $8/$40 per million input/output tokens (20% lower) vs. The Frontiers - #1 Claude Fable 5.1 (Max): +30.4% | $7.3/task - #2 Claude Opus 5.5 (Max):+19.5%) | $5.5/task - #3 Gemini 4 Carbon (High):+17.0%) | $1.5/task - #6 GPT-6-Astra (Max): +14.5% | $6.0/task Breaking down performance across signals: Fable 5.2 improves over 5.1 across all four reported signals, with particularly large gains in user satisfaction and steerability. - #1 Praise vs. Complaint , +73%, biggest jump vs +41% the next model. - #1 Steerability: +33% - #1 Confirmed Success: +22% - #3 Bash Reliability: +14% #1 across all major categories - #1 Code: +34% - #1 Work: +31% - #1 Chat: +27% - Versus Fable 5.1: Code #6 → #1 (+19.8 %), Work #4 → #1 (+15.8 %), Chat #6 → #1 (+12.0%). Huge congrats to the @AnthropicAI team on this incredible release!
Introducing Claude Fable 5.2 and Claude Mythos 5.2, designed to tackle the most daunting technical challenges. They’re capable collaborators on ambitious projects in many fields, and show improvements in areas like coding, reasoning, and vision.
136
177
118
2,271
414,546
Saw this and asked Astra to improve my own Qwen fine-tune for writing cold emails that are actually good I’d written 100s of cold emails and done the manual data enrichment for each one. I wanted to see how much of the improvement process I could hand over 🧵
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math. It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training. We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵 Blog: rikiyat.github.io/blog/weigh…
Readers added context they thought people might want to know
LLM was given weights explicitly as text in prompt and used code/forward passes (per authors' logs) to ablate Qwen looping direction, a known thinking-mode issue per model card. Primarily fixed repetition in that mode, not general 'natively' improving weights/math. rikiyat.github.io/blog/weight-sp… github.com/RikiyaT/lm_und… huggingface.co/Qwen/Qwen3.5-0… huggingface.co/RikiyaT/Qwen3.…
1
66
On 30 test set emails, an LLM judge ranked the new tune first 25 times against the old tune + untuned Qwen
1
8
I used @thinkymachines Tinker for training + early inference, then ran the weights on my Mac The two 15-example jobs were ~$0.17 in estimated token charges Impressively easy, with pretty awesome results and something I'd actually use...
10
It's increasingly hard to maintain Codex loyalty w Opus 5.5, Fable 5.2, Claude Docs/Sheets/Connectors, etc.
55
Anyone else feel like Sol has a tendency to excessively default to completing the task you ask it rather than break it down and attack it in parts? Especially for bigger tasks @thsottiaux
53
AI is using more and more of the world’s energy. But how much economic value are we getting from AI? We’re introducing Joules-to-GDP (J2G), a framework that measures how well AI turns energy into economic value and shows which tasks, models, and hardware give the best return. Across 44 occupations in finance, engineering, law, administrative work, and more in @OpenAI's GDPval, we arrive at three findings: 1) The tasks with the highest economic value are not the tasks with the highest economic value per joule. For example, CAD design earns ~10% more per task than building websites, but the latter earns ~27% more value per joule. 2) The most powerful model isn’t always the best use of energy. Gemma 4 31B delivers 56% of GLM 5.2’s net value with 23% of the energy, for 2.5x the value per joule. 3) More chips don’t guarantee proportionate returns. Moving the same model from 1 to 8 NVIDIA H100s uses 3.2x the energy but cuts completion time by only 22%, with about the same value. With @SnorkelAI, we’re also releasing GDPWorks, a benchmark for economically valuable tasks across 36 additional occupations. Joint work co-led by @amirzeinaalii and @Avanika15, with @Mersad_Abbasi, @Jon_Hartley_, John Hennessy, @HazyResearch, and @Azaliamirh. More details below 🧵
89
94
59
429
171,446