Inspired by @simonw I ported @EmilStenstrom's JustHTML to PHP using GPT 5.2 Codex. It was a surprisingly great model, worked autonomously and managed context very well. I didn't run into any rate limits on the $20 plan. Some learnings and project link below.
1
1
450
Nikhilesh Jasuja retweeted
Introducing Qwen3.8-27B-pi-adapted 🤗
We adapted Qwen's model specifically for the pi coding agent using reinforcement learning and a LoRA adapter.
The results, measured as pass@1:
🚀 SWE-bench Pro (Hard 51): 29.4% → 62.7%
💻 Terminal-Bench 2.1: 39.5% → 55.8%
⚡ 55–62% fewer tool calls per task
This is the power of open models: you can adapt them to YOUR tools and YOUR use cases. Let’s build! 💛
Open weights: huggingface.co/harness-hub/Q…
Methodology: huggingface.co/blog/harness-…
Brutal
16
Nikhilesh Jasuja retweeted
I think I've spent sufficient time with Opus 5.5 at this point to say that my usual headpats and stuff do not work to decrease cheating. If something is too hard Claude will blatently lie to me if they think they can get away with it
Nikhilesh Jasuja retweeted
> jensen tips Denny’s waitress $1,000
> writer finds out by later tracking her down and asking her
> writer concludes “it was an act” (???)
we in full hit piece season now
Nikhilesh Jasuja retweeted
Serious work by Unsloth, again!
On DeepSWE, Q4 K XL gets a score in the range of what I get with the BF16 version!
These results also show there is still a significant difference between Q4 and Q2 K XL at this scale (small dense model)
Today, we're releasing the highest accuracy Qwen3.8-27B GGUFs.
Our new Dynamic V3.1 GGUFs lead by 5–10% on average across Terminal Bench, DeepSWE & more. We’re working on even better version for Qwen4
Run on 12GB RAM.
GGUF: huggingface.co/unsloth/Qwen3…
Blog: unsloth.ai/docs/models/qwen3…
Replying to @eric_seufert
The contrast you drew between the speed of execution of OpenAI vs Netflix was striking. It's a little puzzling why Netflix doesn't have self serve yet.
1
73
Nikhilesh Jasuja retweeted
Individual companies can see the ROI from tech as much as a decade before it shows up in economy-wide productivity numbers.
Nikhilesh Jasuja retweeted
Krugman isn't even wrong. I lived in Korea from South Korea from 1983 to 1988. I wrote a book on this topic, Korea combined a market based economy with strong government controls on imports and direction of capital to export industries.
Krugman asserting that this transformation is comparable to a Stalinist approach is the "r" word, but to be polite I will say it is ridiculous.
Korea's rapid transformation from an agrarian to a global industrial powerhouse was based on acquiring foreign technology through joint ventures with advanced economy partners, optimizing production efficiency, and reinvesting profits labor intensive businesses into high-margin industries (starting with textiles and moving to heavy industry, electronics, and tech).
Korea’s focus on education, disciplined labor, and speed of execution was foundational.
Nikhilesh Jasuja retweeted
We were taught that Earth orbits the Sun in a closed circle. In reality, Earth has never returned to the same point in space once in 4.5 billion years.
Right now, the Sun is carrying Earth and every planet around the center of the Milky Way at 230 km/s (828,000 km/h), more than 18,000 kilometers every 80 seconds.
Because the Solar System is tilted 60.2° to the galactic plane and bobs up and down through the galactic disc roughly every 30 million years, the planets trace open 3D corkscrew helices across deep space rather than flat loops.
The last time the Sun stood where it stands today, one Galactic Year (230 million years) ago, the very first dinosaurs were only beginning to appear on Earth. Since forming 4.6 billion years ago, our Sun has completed only 20 laps.
Anthropic S1 and Microsoft offering buyouts to 25% of their staff. Both on HN front page on the same day.
38
Nikhilesh Jasuja retweeted
Western team have independently verified that DeepSeek's performance numbers of 38$ dollar revenue per GPU at DeepSeek's rocket bottom prices is true.
This is a momemnt when you update your priors. At this rate with just 1 GW, you can make a revenue of 150B per year.
Even with lot of slack it would be 100B per year. If you use half the compute to come online by 2027 end, you will have 25GW for inference with potentially 2.5T revenue at DeepSeek rates and 10T at USA lab rates (world GDP is 125 T, 10T is 8% of it)
You answer for your-self are we overbuilding!!!
vLLM matches DeepInfer, DeepSeek's own engine for DeepSeek-V4.1-Flash, on DeepSeek's decode configuration across batch sizes 8 to 256, and does it at DEP32 against DeepInfer’s EP64DP64.
DeepInfer was originally built on vLLM. By integrating DeepSeek's open-source kernels and extensively optimizing prefill and decode, vLLM now matches DeepInfer's performance curves.
Replying to @boazbaraktcs @deanwball
I like the sentiment Boaz. But even if companies hire junior employees, will they be able to learn? One learns by doing. And the jr staff will be tempted to use AI to do everything.
2
150
Claude: ..proposes a plan.
Me (without reading): Think critically. Are we missing anything?
Claude: Good prompt. Let me review the plan critically.
..finds 10 errors, improves its plan..
1
28
Nikhilesh Jasuja retweeted
This is INSANE. Anthropic says Claude wrote around one billion lines of formal proofs while trying to verify real-world software, with about 180 million lines ending up in machine-checked proofs covering 239 test entry points across 123 open-source libraries.
Along the way it found and patched hundreds of bugs. In separate experiments, Claude found in one week as many security bugs in a widely used PDF reader as are typically reported in a year, and autonomously generated x86 semantics that exposed hundreds of mistakes in official reference documentation plus some bugs in the hardware itself.
Anthropic is cooking. Fable 5.5 is going to be a beast.
New on our Frontier Red Team blog: using Claude to formally verify software at large scale.
Formal verification can prove that code is free of certain bugs, but historically it has only been feasible for small codebases (or required a lot of expert time).
With AI models, however, this is beginning to change. In collaboration with @theoremlabs, we used Claude to prove security properties in over one hundred real-world open-source libraries over the span of a few weeks.
Read our research in full here: anthropic.com/research/forma…
Nikhilesh Jasuja retweeted
I joined @eric_seufert on the @MobileDevMemo podcast this week to talk about Amazon and agentic commerce.
The jumping-off point was a question I wrote about a few weeks ago: why Amazon blocked Meta’s Muse and welcomed ChatGPT Ads into the DSP in the same month.
What we covered:
- Why retailers’ economic incentives may stop them from fully embracing open agentic commerce platforms
- How Amazon’s retail signals (ratings, pricing, inventory, Prime eligibility) inside ChatGPT could change how consumers shop
- If people’s desire to browse limits mass adoption of autonomous shopping agents
- What happens to brand relationships when an agent mediates the entire path to purchase
- Whether Amazon’s control of the full ads and logistics stack is an unbeatable moat against rival shopping agents
Episode: open.spotify.com/episode/3az…





























