Caden Juang
10 followers · 9 following · 501 views
on the atlas · 40
- Design | Apache Kafka1 savers
- The load-bearing vocabulary of Claude5 savers
- Algebraic Effects for the Rest of Us — overreacted1 savers
- Ordinary Abundance20 savers
- Life on the Uncanny Precipice | Naomi Saphra3 savers
- science_halflife - D1-0121 savers
- Retire the Abstractions · Hazy Research4 savers
- Gossip Protocol | Nomad | HashiCorp Developer1 savers
- Introducing FrontierCode | Cognition2 savers
- A Vision of Metascience8 savers
- The Advantages Of Flexible Typing1 savers
- An underleveraged modality in neuroscience | Abhi Upadhyay13 savers
- Obama Presidential Center | From "Me" to "We": A New Monument for Democracy + Interviews - YouTube1 savers
- Dune: Imperium – Uprising | Board Game | BoardGameGeek1 savers
- Cosmic Encounter | Board Game | BoardGameGeek1 savers
- Rohin Shah on what it's really like to run AGI safety at Google DeepMind (and where I disagree with 'doomers') | 80,000 Hours1 savers
- How to Buy Cheap Claude Tokens in China - by Zilan Qian8 savers
- What I've been thinking about this weekend - More open questions, intelligence vs power, the problem of verification in science, the parallel discovery of Darwinism1 savers
- Automated Weak-to-Strong Researcher17 savers
- The speed of science - Works in Progress3 savers
- Native Instant Space Switching on MacOS1 savers
- Tiny TPU10 savers
- LoRA Without Regret - Thinking Machines Lab37 savers
- Understanding UMAP9 savers
- [2505.10831] Creating General User Models from Computer Use2 savers
- NAP — Next-Action Prediction1 savers
- Model Runner V2: A Modular and Faster Core for vLLM | vLLM Blog2 savers
- AI Inference Benchmarks | InferenceX by SemiAnalysis1 savers
- GPU Glossary11 savers
- Mamba: The Hard Way9 savers
- 0xPARC1 savers
- DP, DPA and SGLang DP Router — SGLang1 savers
- [2602.04770] Generative Modeling via Drifting2 savers
- The lump of cognition fallacy - Andy Masley2 savers
- [2601.11516] Building Production-Ready Probes For Gemini1 savers
- Inter-process communication in Linux: Shared storage | Opensource.com1 savers
- Gruvbox Rainbow Preset | Starship1 savers
- Pain is not the unit of Effort38 savers
- Julia10 savers
- The Basics10 savers
highlights · 36
How do we iterate on rubric quality?
Introducing FrontierCode | Cognitionfor misalignment or loss of control, the threat model is tied more around internal deployment rather than external deployment
Rohin Shah on what it's really like to run AGI safety at Google DeepMind (and where I disagree with 'doomers') | 80,000 HoursSo if you tie evals to, and require them to be pre-deployment, then that is providing a pretty strong incentive to make those evals as fast as possible to run and get them done as quickly as possible, which is maybe not really the incentive you want to give.
Rohin Shah on what it's really like to run AGI safety at Google DeepMind (and where I disagree with 'doomers') | 80,000 HoursAs technology advances, it gets difficult to make new useful discoveries and to advance it further. That’s the first one. The second one is, as technology advances, we have better tools to do science and to make new discoveries. The third one is, as technology advances and the economy grows, we can support a larger number of people who will do the research and have the ideas and advance science.
Rohin Shah on what it's really like to run AGI safety at Google DeepMind (and where I disagree with 'doomers') | 80,000 HoursA screenshot circulated in a developer WeChat group joking about the supply chain to bypass Anthropic’s KYC; originally in Chinese (up), translation added by the author on the bottom
How to Buy Cheap Claude Tokens in China - by Zilan QianAgents travel to lower-income countries in Africa or Latin America to recruit real individuals willing to complete in-person verification.4 The Worldcoin black market offered a documented precedent, with iris scans harvested from KYC merchants in Cambodia and Kenya, sold for under $30.
How to Buy Cheap Claude Tokens in China - by Zilan QianWhile this setup sounds functionally identical to legitimate Western API aggregators like OpenRouter,
How to Buy Cheap Claude Tokens in China - by Zilan QianTransfer stations, conversely, are built explicitly for evasion, routing data through unaccountable middlemen.
How to Buy Cheap Claude Tokens in China - by Zilan Qianmaking Claude the first major consumer AI platform to implement this level of identity checking.
How to Buy Cheap Claude Tokens in China - by Zilan Qianthe history of human science shows that the verification loop for theories can be on the order of decades and centuries, and even then experiments do not definitely rule out alternatives
What I've been thinking about this weekend - More open questions, intelligence vs power, the problem of verification in science, the parallel discovery of DarwinismWithout prescribed steps, AARs behave more like expert researchers: they design fast, cheap experiments to test the hypothesis first and only commit to full training once the premise holds.
Automated Weak-to-Strong Researcheruns gradient-free Evolution Strategy optimization of LoRA parameters, using unsupervised swap-consistency as the fitness signal — perturbations are rewarded for producing predictions that are both confident and symmetric (p(A>B) ≈ 1 − p(B>A)). Aggregates 16 seeds via swap-consistent inference and trimmed-mean in logit space.
Automated Weak-to-Strong ResearcherGeneralization to held-out datasets.
Automated Weak-to-Strong Researcherour AAR usually runs extensive ablation studies during hill-climbing
Automated Weak-to-Strong ResearcherWe find the directed setting effectively prevents entropy collapse. In comparison, in the undirected setting, AARs concentrate on only a few directions like self-training and collapse quickly.
Automated Weak-to-Strong ResearcherLines of raw Python code Lines of Claude-generated pseudocod
Automated Weak-to-Strong Researchertwo authors spent 7 days tuning four representative prior methods
Automated Weak-to-Strong Researcherdivision of labour into science through norms and institutional policy
The speed of science - Works in Progressif peer review was a speciality
The speed of science - Works in ProgressAnd as databases are created and research is contracted and disseminated openly, there would be an ever-growing number of scientists who would become contributors to an expanding universe of scientific knowledge.
The speed of science - Works in ProgressPeer review would operate openly
The speed of science - Works in Progressdeveloping appropriate measures to record data, and testing the assumptions of their research – as the meta-scientist Anne Scheel and others have suggested. Scientists would use programs such as DeclareDesign to simulate data, and test and refine their methods.
The speed of science - Works in ProgressFrom the year 2000 onwards, scientists who have conducted clinical trials have been required to declare the methods they will use to analyse their data before they even collect it.
The speed of science - Works in ProgressJOSS is actually built upon the platform of GitHub: software developers submit their software together with working papers describing them to the journal via GitHub
The speed of science - Works in ProgressBut more broadly, individual researchers no longer have to rediscover these protein structures over and over again if they have already been shared publicly, if they want to use that knowledge for other purposes.
The speed of science - Works in Progressis all of science like this?
The speed of science - Works in ProgressWhen non-scientists, who want to apply scientific knowledge elsewhere, read research,
The speed of science - Works in ProgressBut unfortunately peer review often fails to accomplish this, as the scientist Stuart Ritchie has described in his book Science Fictions.
The speed of science - Works in Progressprior research that has been disputed, corrected or retracted.
The speed of science - Works in ProgressWhen scientists reference previous research, they cite entire papers, not specific results or values from them.
The speed of science - Works in Progresspapers – static documents
The speed of science - Works in ProgressLouis Anslow who runs the Pessimists Archive coined a useful simple extension of this idea: the lump of cognition fallacy: the idea that there is a fixed amount of thinking to do. I see this come up a lot in AI discourse. Like the lump of labor fallacy, the reason this is wrong is simple: thinking often leads to more things to think about.
The lump of cognition fallacy - Andy MasleyThe obvious challenge in using this IPC mechanism is that a race condition might arise: the producer and the consumer might access the file at exactly the same time, thereby making the outcome indeterminate. To avoid a race condition, the file must be locked in a way that prevents a conflict between a write operation and any another operation, whether a read or a write. The locking API in the standard system library can be summarized as follows: A producer should gain an exclusive lock on the file before writing to the file. An exclusive lock can be held by one process at most, which rules out…
Inter-process communication in Linux: Shared storage | Opensource.comIn the words of author Michael P. Farrell, “extraordinary creativity is often the result of successful collaboration among peers who develop an intense friendship and work together on similar projects for an extended period of time”. History makes a compelling case for this.
The BasicsIn the words of author Michael P. Farrell, “extraordinary creativity is often the result of successful collaboration among peers who develop an intense friendship and work together on similar projects for an extended period of time”.
Pain is not the unit of EffortIf anything they found escape and respite in working on their education, and having hope and a direction to work towards increased their subjective well-being.
Pain is not the unit of Effort