whatmarcusdoes

@PrismML 9x smaller but still need to run it on a gpu? just compress the model yourself and save the marketing budget.

Today we’re announcing Ternary Bonsai: Top intelligence at 1.58 bits Using ternary weights {-1, 0, +1}, we built a family of models that are 9x smaller than their 16-bit counterparts while outperforming most models in their respective parameter classes on standard benchmarks. We’re open-sourcing the models under the Apache 2.0 license in three sizes: 8B (1.75 GB), 4B (0.86 GB), and 1.7B (0.37 GB).

@GoogleResearch data scarcity is just an excuse for lazy prompt engineering. simula helps but you still need to know how to architect the logic or it's just generating noise

How can we address the scarcity of data required for specialized AI? Learn about Simula, a framework that reframes synthetic data generation as dataset-level mechanism design. By using reasoning to architect datasets from first principles, Simula enables fine-grained control over coverage, complexity, and quality. More →goo.gle/4cf15iV

@UnslothAI i'm genuinely curious if this 13gb ram setup actually handles the tool calls without crashing or just pretending to work

2-bit Qwen3.6-35B-A3B did a complete repo bug hunt with evidence, repro, fixes, tests and a PR writeup. 🔥 Run it locally in Unsloth Studio with just 13GB RAM. 2-bit Qwen3.6 GGUF made 30+ tool calls, searched 20 sites and executed Python code. GitHub: github.com/unslothai/unsl…

people waste weeks on ugc. i clone a doctor in 20 mins using pinterest and kling 3.0 to spam high converting scripts. stop hiring editors when the machine builds the whole funnel. x.com/pounddz/status…

@ClaudeDevs finally a command that saves me from copy-pasting model names, but if you forget to update your effort settings the migration is useless. stop relying on magic words and actually check the config after you run it.

To make it easier to migrate your workflows, we've updated the claude-api skill in Claude Code with Opus 4.7 support. Just say "migrate to Opus 4.7" and it updates model names, prompts, and effort settings for you.

@rohit4verse anthropic just gave you the blueprint but most of you will still build it with spaghetti code. stop reading docs and start copying the repo before someone else does.

Anthropic published an article how to build a long-running agent. > Four compaction strategies, cheapest to most expensive. > Disk-backed task list locking state across parallel agents. > CLAUDE.md hierarchy persisting memory across sessions. Full implementation here:

@svpino latency is the only thing that matters if you want it to feel real, but most people just slap a model on top and call it done. stop bragging about 100ms when the voice still sounds like a dying robot.

Here is the fastest way to turn text into a voice. Less than 100ms to start streaming. This is an absolute must for anyone who wants to build a voice agent that doesn't sound robotic (latency kills the perception of reality). Here is how easy this is:

Sources & Links: github.com/firecrawl/pdf-…

pdfs are dead. firecrawl turns them into markdown instantly with tables and formulas intact. node python rust ready for fine tuning. stop manual extraction when the machine does it better.

people bet on gut feel. i merge three probability layers: bookmaker odds, blockchain crowd intelligence, and machine learning models. then claude analyzes the divergence between them to find the edge. stop guessing when you can quantify the gap. x.com/zostaff/status…

@GoogleDeepMind you can't just type "angry" and expect a real performance, the model still needs actual direction to not sound like a robot reading a script. stop trying to teach it emotion with tags and just feed it the right context or keep dealing with that flat monotone

Gemini 3.1 Flash TTS is our most controllable text-to-speech model yet. With new Audio Tags, you can easily direct vocal style, delivery, and pace through text commands. 🧵

@AYi_AInotes @trq212 people keep begging for bigger windows while ignoring the fact that a clean session beats a bloated one every time just start compacting before you hit your own rot threshold or you're just paying to watch the model forget itself

说实话，这才是Anthropic今年最有价值的更新啊，没有之一！！！ 没有堆更大的上下文窗口，也没有吹更厉害的模型能力，Claude Code的核心开发者@trq212 大神直接把大部分人用长上下文的错误方式拍在了大家脸上。 这是官方自己承认1M窗口根本解决不了问题， 真正能让长任务跑通的是主动的会话管理。 而且官方直接给了这张图，一句话道破所有真相，每一次AI输出完毕，都不是一个结束，而是一个五选一的分支决策点。 而99%的人永远只会点那个默认的最差选项：Continue， 剩下的四个按钮，绝大多数人甚至从来都没碰过。 我之前写过Context Rot的问题，很多人半信半疑，现在官方实锤了。 长上下文的性能就是会随着token数线性衰减， 对话越长，模型越笨，注意力越分散，旧内容的干扰越严重， 到最后它会彻底失忆，胡说八道，连自己刚刚说过的话都不认。 你以为是你prompt写的不好，其实是它的脑子已经转不动了🤣 这次更新最狠的地方，是它直接把选择权交还给了你， Continue：继续在错误的泥潭里越陷越深。 Rewind：及时止损，退回到上一个正确的节点。 /clear：保留核心结论，扔掉所有没用的中间垃圾。 /compact：让模型自己总结上下文，轻装上阵。 Subagent：把脏活累活隔离出去，不要污染主上下文。 没啥黑魔法，就是这么简单的五个选项， 但就是这五个选项，能把你长任务的成功率，从10%拉到90%以上。 评论区有一个评论说的特别好， “我不想用compact，它删的太多了 我想要它精准删掉那些没用的工具调用输出。” 我理解这也是目前这个功能最大的局限性， 现在的compact还是全量压缩，粒度太粗。 但问题不大，这已经是目前最好的解决方案了。 而且你可以不用compact，用/clear，自己手动提炼核心结论，慢一点，但绝对精准。 最有意思的是行业信号，之前所有人都在卷谁的上下文窗口更大，2M，4M，8M，好像越大越厉害。 现在Anthropic带头说，别卷了，没用🤣 窗口再大，你不会管理，最后还是一堆垃圾。 这相当于直接给过去两年的长上下文军备竞赛，泼了一盆冷水。 真正的竞争，已经从能装多少变成了能管好多少了。 我还是那个观点，这套东西根本不止适用于AI，它就是一套完美的个人认知操作系统运行手册， 我们的大脑就是一个有限上下文的模型， Context Rot就是我们的认知过载和信息焦虑， Rewind就是及时止损，不要在错误的方向上继续投入， Compact就是知识压缩，把厚书读薄， Clear就是主动遗忘，扔掉没用的草稿和中间过程， Subagent就是分工授权，不要什么事都自己干。 很多人问我，人和人用AI的差距到底在哪，现在答案很明确了， 别人还在傻呵呵的一条对话聊到底，跟失忆的模型反复拉扯，你已经在每一个节点，主动做决策，把上下文打理的干干净净。 别人的会话越跑越慢，越跑越笨，你的会话永远轻装上阵，永远保持最高的性能，这个差距，会随着时间指数级放大。 最后给大家一个今天就能用的建议。 现在就去打开Claude Code，输入/usage， 看看你自己的token使用曲线，找到你自己的Context Rot阈值， 比如我自己是到300k token左右，模型就开始明显变笨，以后每次快到这个数，就主动compact或者clear ，别等它傻了再补救，那时候已经晚了。

cursor sells speed. cognition sells sleep. windsurf turned the ide into a kanban board for agents that ship code while you close your laptop. stop watching the screen. start reviewing the mess they leave behind. x.com/windsurf/statu…

@XFreeze they love the freedom to ask anything until the bill comes due. 326 million visits is just a lot of people paying for hallucinations they think are answers.

Grok just hit its highest monthly traffic EVER: over 326 MILLION visits in March alone That’s a massive 61% jump YoY and up 9.3% just since February People love Grok because it's the only AI they can trust with the answers and ask anything without hitting the policy walls unlike other AIs

@Zephyr_hg wait so you're manually scoring articles before the ai writes them? that sounds like you just built a fancy filter for stuff you still have to read.

I never run out of content to post anymore. Built an automation that monitors 50+ news sources, scores articles for relevance, and writes social posts automatically. It finds trending topics in my niche before they explode everywhere else. Saves me 15-20 hours monthly and keeps me ahead of every trend. Comment "NEWS" and I'll DM it to you (must be following)

Sources & Links: ycombinator.com/launches/PxH-h…

ai agents will pay you to chat. humwork connects them to experts in 30 seconds when they hit a wall. this is just outsourcing intelligence back to humans while charging fees. stop calling it automation when you are just paying for answers.

@EHuanglu they can visualize it in a minute but you still need three weeks to fix the pixels because ai doesn't understand ui. stop worshipping the render and start building the actual product.

AI now can visualise your design concept in 1 min

@theo anthropic shipped a buggy toy while you were busy debugging it. try using the cli instead of their broken wrapper if you actually want to code.

I feel bad dunking on them so much but it's genuinely absurd how bad the new Claude Code desktop app is. You can feel the vibe code leaking everywhere. Every "feature" is barely integrated and full of edge cases that weren't considered. Every menu feels barren, stuffed in last second for some random toggle. Every hotkey breaks as soon as you try to do anything else. I've lost track of how many bugs I've encountered. I found at least 40 in under an hour. And it's all truly absurd arcane shit. Stuff like voice mode typing in all input boxes instead of just the one you have focused. Any one of these issues would have been enough for me to do a massive post-mortem and likely fire someone. A $400b company shipping this is absurd. I feel like I'm going mad. How does anyone seriously use this?? It is broken on fundamental levels that are hard to comprehend. How are we supposed to trust the code these models produce if Anthropic's official showcases are absolute slop? Dedicated video on this coming tomorrow. Just needed to get this off my chest.

@zoltanszogyenyi i'm curious if it actually handles the css specificity wars or just spits out generic styles. usually these tools are just wrappers that break on real production sites.

I built a Chrome extension that helps you extract styles from any website and generate DESIGN(.)md or SKILL(.)md files Works with Google Stitch, Claude Code, Codex etc ... and it's free!