Nathaniel Bechhofer retweeted
Run Inkling 2 Mini, the strongest 26.4B model that runs on 16GB RAM.
Thinking Machines’ Apache 2.0 MoE model uses a looped transformer arch & scores 48.2% on DeepSWE v1.1, the best for size.
Run on Unsloth, llama.cpp
GGUF: huggingface.co/unsloth/Inkli…
Guide: unsloth.ai/docs/models/inkli…
Some of you thought that Inkling 2 Small was too big at 594B parameters. We heard you: Inkling 2 Mini is a looped transformer with just 26B parameters, small enough to run on your laptop, powerful enough for agentic work with a score of 20.1 on AA Index.
thinkingmachines.ai/news/ink…
Nathaniel Bechhofer retweeted
a good day to repost:
> if a tool can solve millennium prize problems and be this dumb on 1000 other things, that reshapes my understanding of the territory more than it reshapes my judgment of the vehicle
Replying to @aryaman2020
my post is precisely a result of seeing the conversations about that
(1) part of it is: if a tool can solve millennium prize problems and be this dumb on 1000 other things, that reshapes my understanding of the territory more than it reshapes my judgment of the vehicle
it's becoming clear that, well, compute-vs-humans are mutually jagged
(2) and part of it is: the brute force vs. generalization ratio in the current paradigm has hugely missed my (rather low) expectations, e.g. "sorry but my AI that solved Navier Stokes needs help so it can write English or explain itself, after seeing the whole Web 10x over"
(3) and part of it is about a different philosophy of value/work: [a] whatever AI makes cheap (down to and including em dashes) reliably gets displaced in value by other stuff, [b] there's no short reward function that re-creates all value or information, most of it is inherently distributed and can't be pre-loaded
Nathaniel Bechhofer retweeted
Meet our first interactive explainer on gemma․ai: How LLMs work (with Gemma 4)
We are expanding gemma.ai with content for everyone: visual guides, interactive guides, hands-on guides, and more!
This first explainer uses the Gemma 4 tokenizer directly in your browser to help you learn about how LLMs work.
We love your feedback! What would you like to see next on gemma.ai?
Nathaniel Bechhofer retweeted
Harnesses used to be a crutch for weak LLMs. In 2026 and onward, with stronger models, harnesses now offer a mechanism for composing an LLM’s existing and new capabilities to generalize well beyond their training environments.
In our new paper, LM Harnesses are Compositional Generalizers, we explore how various harness abstractions enable LM systems to generalize over task length, task strategies, recursive task depth, and more through composing existing LLM capabilities. These abstractions provide a means for more data-efficient training, especially over long tasks, by keeping trajectories locally in-distribution (LID) — i.e., shaping all individual LLM calls to be in-distribution.
For tasks that take a very long time to roll out, this implies that we can even perform surrogate harness training, meaning we train an individual LLM call instead of doing the very long harness rollout, using purely synthetic tasks, with fake tool calls + sub-agents!
A common idea floating around is that LLMs will absorb harnesses and make them obsolete. We disagree, and argue that we haven’t sufficiently explored what harnesses can provide.
The full paper, logs, codebase, and data are available: alphaxiv.org/abs/2611.harnes…
w/ @SOURADIPCHAKR18 @lateinteraction
Nathaniel Bechhofer retweeted
New post. Rizzo and Whitman’s Escaping Paternalism: Rationality, Behavioral Economics, and Public Policy — a review blog.jcx.au/posts/rizzo-and-…
Nathaniel Bechhofer retweeted
Improved cash transfer tranching and timing can better suit poor households’ unique financial needs. Just Accepted new paper by Carolina Kansikas, Anandi Mani, and Paul Niehaus @PaulFNiehaus zurl.co/EJuoA
Nathaniel Bechhofer retweeted
Read this article. The author literally brings zero receipts to show that surprise billing is raising spending. Best evidence we have is that arbitrated cases get big payments. The @karmaroge in the @TheAtlantic was...not fantastic.
Nathaniel Bechhofer retweeted
Folks I am BEGGING you. Stop reporting youth unemployment rate trends without looking at employment and participation. This a huge source of confusion and its late to be still making this mistake.
Unemployment rates are an increasingly bad signal about the health of the youth labor market.
The chart in the thread below, showing a steep rise in unemployment rates for young master’s degree holders, is a great example.
Employment rates, the percent of all young master’s degree holders that are employed, is flat since 2019. A very different story!
You can also see, as @ModeledBehavior and I have pointed out, those without a BA have seen employment rates fall as much or worse than those with a BA.
Nathaniel Bechhofer retweeted
Out today!
econometricsociety.org/publi…
Nathaniel Bechhofer retweeted
🎉 New CSES data!
Module 6 Third Advance Release: 55,207 respondents, 29 election studies, 27 polities.
PLUS: Our new Data Playground! 📊
Thanks to our global collaborators and CSES Secretariat at @VU_CGD & @gesis_org!
🔗 cses.org/2026/11/17/cses-ann…
#CSES #ElectionStudies
Nathaniel Bechhofer retweeted
This paper is the first to analyze immigrants’ fiscal effects over the 20th & 21st centuries, showing that immigrant surpluses have surged as deficits have exploded. Since 1900, immigrants reduced deficits by $14.6 trillion (excluding interest savings).
cato.org/policy-analysis/imm…
Nathaniel Bechhofer retweeted
the thing that everyone forgets is that COMPUTERS ARE FAST
waiting SECONDS for things to load is NOT NORMAL
Nathaniel Bechhofer retweeted
⚠️ Numerical issues can go on silently for decades like in the softmax/CE! Similar issues are present in other widely-used divergences like the KL and JSD
I just posted this paper overviewing what they are and how to fix them:
📚 arxiv.org/abs/2611.18511
Nathaniel Bechhofer retweeted
A nice way to promote human collaboration in the age of AI
Ok, this may be my favourite teaching thing ever: for a class in which students write a paper, they were divided into small groups, repeatedly, to share their work with each other and get feedback. But then: after class, they filled out a survey with who helped them and how. They could each nominate up to three people for bonus participation points.
And the responses are so heartwarming! First, it's great to see some people really stepping up and helping a lot of their classmates. Second, the answers are creative and funny, like the one about someone who helped out by not understanding (thereby forcing a rethink in how they explained their work). Third, students who didn't know each other exchanged contact info to stay in touch later, which I count as a win. Worked out really well, would do again.
Nathaniel Bechhofer retweeted
in this post the Claude-Code dev team boast that 90% of claude-code code is written and verified by agents, without a human review. and then they provide some generic tips about what are good tests etc.
but my take is that the real reason they can pull it off is not their amazing auto-testing skills, but the fact that claude-code is essentially a very simple piece of software, they don't have almost any efficiency concerns, no competitors, very little potential of actually harmful bugs or data corruption due to claude code itself, and that many bugs just got swallowed and compensated for by the underlying LLM being so freaking good.
Claude writes better code when it can verify its work. Checks that used to be expensive, like tests, benchmarks, and end-to-end runs, are now cheap.
Here’s how the Claude Code team thinks about verification:
claude.dev/blog/how-the-clau…
codeberg.org was suspended at the .org registry an hour ago, lmao...oh my lord please don't put your projects on here w/o a mirror somewhere else
Nathaniel Bechhofer retweeted
I recently became aware of how confused even many academics are about Pangram. Please, please, look at the audits that have been done before you deliberately sow doubt.
(link in thread)
Nathaniel Bechhofer retweeted
Sagebrush is what I always wanted to write when I started SageMath. AI made it possible.
sagebrush.space/blog/sagebru…
Nathaniel Bechhofer retweeted
New working paper with the brilliant Michal Kolesár:
Shrink Responsibly: Empirical Bayes in Economics
We review the ins and outs of eB "shrinkage," as applied in econ settings, and illustrate them with an application to public school value-added
Some takeaways:
- Conventional normal-normal shrinkage, perhaps with an adjustment for precision dependence, can work pretty well -- even when misspecified (like OLS!)
- Plenty of interesting questions should not be (but often are) answered with shrinkage; we give alternative estimators for some cases
- Elementary and middle schools in Charlotte-Mecklenberg vary widely in value-added, which appears causal. More popular schools are not reliably better, but better schools seem to have higher-achieving peers
Link below!



































