deva

e/acc

[deva @devaxsha]

DECIDED TO REVAMP THE TOKENLEAK WEBSITE

[思维怪怪 @0xLogicrw]

谷歌 TPU 软件工程师 Patrick Toulme 指出，外界对 GLM 5.2 靠蒸馏追平 Opus 的说法存在误解。大模型在智能体编码任务上的训练难点在于「零梯度困境」，即模型早期若无法产生正确运行路径，强化学习便无法获得梯度信号来启动参数更新。蒸馏 Claude 或 GPT-5.5 的作用，仅仅是在冷启动阶段提供种子解答以绕过零梯度困境。

一旦模型跨过冷启动门槛，后续的性能爬升将不再依赖蒸馏，而是完全依靠强化学习的爬山算法进行自我演化。Toulme 强调，GLM 5.2 已经具备独立产生成功路径的能力，完全可以通过强化学习自主迭代到更高级别，彻底摆脱对美国大模型的依赖。

[deva @devaxsha]

important context: Fugu is an orchestrator.

if you read the blog post, it’s basically a multi-agent system trained to route and coordinate a pool of other llms.

that’s collective intelligence boosting the score, which probably means it’s calling the same frontier models it’s being compared against, like gpt-5.5 and opus.

[Sakana AI @SakanaAILabs]

Introducing Sakana Fugu: A full multi-agent orchestration system accessible via a single model API.

Our ‘Fugu Ultra’ model matches the performance of Fable and Mythos, delivering frontier capability without the risk of export controls.

Try it: https://t.co/hhO6qTawgb 🐡

[deva @devaxsha]

mythos class model btw

[Sakana AI @SakanaAILabs]

Introducing Sakana Fugu: A full multi-agent orchestration system accessible via a single model API.

Our ‘Fugu Ultra’ model matches the performance of Fable and Mythos, delivering frontier capability without the risk of export controls.

Try it: https://t.co/hhO6qTawgb 🐡

[deva @devaxsha]

literal large language model

[Merriam-Webster @MerriamWebster]

Our new LLM.

Available now.

[deva @devaxsha]

claude fable being PURPOSELY bad in frontier llm development for other folks to not catch up is so anthropic coded

[deva @devaxsha]

here’s the research blog

https://t.co/mXzA494tSx

[deva @devaxsha]

> apple has shipped a 20B parameter on-device model.

> you can't fit 20B params in RAM at reasonable precision, so they improved the architecture.

> a small model predicts which experts to load from NAND to RAM per query. unlike typical MoE, experts don't switch every token.

[Josh Elman @joshelman]

The technical details behind Apple Foundation Models are worth a read:

https://t.co/RI5qV9eKBU

[deva @devaxsha]

if you're still prompting coding agents directly you're falling behind. you need an agent that infers the task from your cursor hovering over a file for 12 seconds and then over engineers it

[rahul @0interestrates]

if you’re still writing loops that prompt coding agents you’re falling behind. you need to build a meta agent that infers what loops you would have wanted based on your vibe and then write those loops

[deva @devaxsha]

i’m just glad iOS 27 has many performance improvements

[deva @devaxsha]

@pbicho96 underrated thanks for the share

[deva @devaxsha]

this is the kind of AI infra update that actually matters lol

huawei just open sourced KVarN: 3-5x KV cache compression, plugs into vLLM with one flag, and claims speedups instead of the usual quantization tax

If it holds up, longer-context local LLMs get a lot cheaper.

need to start experimenting with it

[deva @devaxsha]

ok so apparently designing loops not prompts is the new thing

a useful coding loop is:

prompt agent → inspect diff → run verifier → decide continue/retry/stop → save output

> the loop is plumbing.
> the skill inside the loop is the asset.

vague loops burn tokens. verified loops compound.

[deva @devaxsha]

@andrelandgraf @rauchg This is so cool Andre! Love this

[deva @devaxsha]

@TheIshanGoswami built tokenleak and timemachinesdk !!!

https://t.co/yPD9PRKOIa

[deva @devaxsha]

Just release v2 of Tokenleak, with a total overhaul using opentui and Solid.js

Now with cursor integration too

monitor your tokens with an even better interface!

[deva @devaxsha]

@sdand sick work! curious to know how you’ve scraped your twitter home feed. Is it via the official api?

[deva @devaxsha]

@NousResearch do they have browser use tools

[deva @devaxsha]

@donnfelker just reverse prompting, give a brief prompt to chatgpt/gemini to give you a prompt for generating an image that does xyz and make it as detailed as possible and many a times it’d make an amazing prompt

[deva @devaxsha]

gpt 5.5 high been oneshotting a lot of whatever i throw at it

im struggling to get used to this paradigm shift where I don’t double check its output

[Frank @jedisct1]

Breaking: The results of Mythos’s audit of the Rust standard library have been leaked https://t.co/lst0VesfrW