Wal

#RedHatAccelerator, RHCA II, CK{A|D|S}, maintaining curated CK{A|S} repos My opinions are my own.

[codila @0xCodila]

Anthropic just released 1-hour course to master 100% of Claude Code:

"Your old prompts making Claude worse - our engineers made 1 prompt cuts costs 90% + improves quality"

0% → 0:27 - Anthropic prompts to master Opus 5.5
25% → 20:21 - /harness for Managed Agents
50% → 29:11 - /loops + /self-improving agent teams
75% → 47:16 - Andrej Karpathy's method
100% → 1:02:15 - 24/7 self-improving graphs on max

This 1-hour course by Anthropic senior engineer will teach you more of using Claude than 100 YouTube video guides...

Learn Claude from Anthropic, master it - then read article on how to reverse-engineer anything with Opus 5.5:

[codila @0xCodila]

REA is most dangerous thing on the internet right now...

It lets your AI agent reverse engineer any app, down to the binary

If you use it correctly, you can rebuild any feature you pay for, for FREE

In this article, I show you how: https://t.co/3P1FNXiXzl

[Not Jerome Powell @alifarhat79]

Who did this shit lmao

[Hisham Al-Barakani @Heshamyemr]

📍هل اعتقدت يوماً أن الذكاء الاصطناعي يمكن أن يكون "فقيهاً" موثوقاً لا يهلوس ولا يؤلف الأحاديث؟ 🤔

أخضعتُ تطبيق "ق" (Qaf AI) العربي لأقسى 5 اختبارات تعجيزية لكسر نموذجه (فخاخ فقهية، قيود لغوية صارمة، وبرمجة معقدة).. والنتيجة كانت صدمة حقيقية: نجاح مبهر بنسبة 100% خالي من أي تأليف! 🤯🤖

نحن لسنا أمام مجرد "شات" عادي، نحن أمام ثورة تقنية حقيقية دمجت بين أحدث نماذج الذكاء الاصطناعي المفتوحة وبين أصالة التراث الإسلامي. إليك لماذا يجب أن يكون هذا التطبيق في هاتفك الآن:

🛡️ وداعاً للهلوسة الدينية: التطبيق مبني على بنية تحتية جبارة (RAG) تم فيها فهرسة ملايين الصفحات من أمهات الكتب. إذا سألته عن حديث مكذوب، سيكشف لك الفخ فوراً ولن يخترع له سنداً كما تفعل النماذج العالمية.

⚖️ أمانة علمية تصدمك: التطبيق مبرمج بأخلاقيات العالم الشرعي! في النوازل الشخصية المعقدة (مثل طلاق الغضب)، يمتنع بصرامة عن إطلاق الفتاوى العشوائية، ويوجهك فوراً للقضاء ودار الإفتاء، مع استعراض الخلاف الفقهي بحيادية تامة.

🧠 عبقرية لغوية وبرمجية: يتجاوز كونه موسوعة شرعية؛ فهو قادر على "إعراب" أصعب الجمل النحوية، الالتزام التام بقيود عدد الكلمات والحروف الممنوعة، بل وحتى كتابة أكواد (Python) معقدة وجدولتها باحترافية عالية.

📲تطبيق "ق" هو معجزة تقنية صُممت لتفهم العقل العربي وتحمي هويته.

هو باحثك الأكاديمي، ومستشارك المؤتمن الذي تضعه في جيبك. 🚀📱

لا تكتفِ بالذكاء الاصطناعي العشوائي، جرب الأقوى والأكثر أماناً الآن:

🔗 رابط التطبيق من متجر أندرويد 📲:
https://t.co/4u9LOml8uK

🔗 رابط التطبيق من متجر آيفون 📲:

https://t.co/D28hUln8FC

@Abdella6if

#الذكاء_الاصطناعي
#تطبيقات
#تقنية
#ChatGPT

[xsmryze @xsmryze]

THIS IS ACTUAL F*CKING GOLD

Anthropic just shared one of the best tricks for Claude Code.

The idea: run Sonnet, Haiku and Opus 5.5 together, each one doing only the job it's best at.

Opus-level judgment on every big call, without paying Opus rates for every routine turn.

You just launch Claude Code like this:

claude --model sonnet --advisor opus

here's the split:

→ Sonnet 5.5 runs the session and writes the code
→ Haiku 5.5 explores files and docs in parallel
→ Opus 5.5 stays on call as advisor for the big calls
→ you stop spending Opus on work that doesn't need it

Opus only steps in at three moments:

→ before a plan locks
→ when the same error shows up twice
→ before calling it done

Sonnet builds. Haiku explores. Opus reviews.

JEV engineering is the same move one layer down: the forks that need no thinker (which file, which tool, retry or stop) go to JEV in milliseconds, and the big models only see the ones that split

the full tree
> Sonnet 5.5 runs the main session and edits code
> search maps the repo on Haiku 5.5
> grep finds call sites on Haiku 5.5
> docs pulls the API reference on Haiku 5.5
> Opus 5.5 on call as advisor, reads the whole transcript when consulted
> JEV routes micro-forks in milliseconds

paste the tree and this prompt into Claude Code ↓

"Rebuild my Claude Code setup around this tree:
1. Run the main session on Sonnet 5.5 with Opus 5.5 as advisor.
2. Create three read-only Haiku 5.5 subagents: search (maps the repo), grep (finds call sites), docs (pulls the API reference).
3. Add one rule to CLAUDE.md: consult the advisor before locking a plan, after the same error twice, and before calling a task done.
4. Route fast deterministic forks (tool choice, path selection, retries) through JEV.
Show me every change as a diff first. No edits until I say go."

↳ https://t.co/bsYb2K5H9G

save it, you'll use this every single day.

[مجلة ميم.. مِرآتنا @Meemmag]

عضو مجلس الشيوخ المنتخب، عبد الرحمن السيد، يلقن نائب الرئيس الأمريكي درسا: من المضحك حقاً أن جي دي فانس وصفني اليوم بـ “الشخص الحقير“؛ أنا أشبَه بـ “حيوان الوولفرين” الشرس، فليحتفظ هو بلقبه “باك آي” (Buckeye) وليذهب به إلى الجحيم!

[Sprytix @Sprytixl]

THIS IS ACTUAL F*CKING GOLD

Anthropic just shared one of the best tricks for Claude Code.

The idea: stop configuring Claude Code by hand. One prompt, 44 seconds, and your whole .claude/ folder writes itself.

7 files, 162 lines, 2 cents. you typed one line and wrote zero of them.

You just launch Claude Code like this:

claude --model opus "set up this repo: ..."

here's the split:
→ 3 Haiku 5.5 agents scan the repo before a single file gets written
→ one reads the stack, one reads the npm scripts, one hunts for risky paths
→ Opus 5.5 takes those three summaries and writes the actual config
→ 7 files land: CLAUDE.md, 3 agents, permissions, a format hook and a /ship skill

the part everyone gets wrong: they write CLAUDE.md from memory.

a config written without reading the repo is a guess, and you live with that guess for months.

here the risk scan came back with 5 dangerous paths and those went straight into the permissions file, so the agent is blocked from them before it ever runs once.

the scripts scan found 5 npm scripts, and the agents got wired to the ones that actually exist in this repo instead of generic placeholders.

one prompt from you, 7 files from it, 0 lines written by hand.

Haiku reads. Opus writes. You commit.

JEV engineering is the same move one layer down: 1,867 forks this run - which file, which agent, allow or ask - resolved in code at 0.97 sharpness, and the split ones (0.59 on "allow or ask") come back to you instead of being guessed

the full tree
▸ you paste one prompt describing the setup you want
▸ 3 Haiku 5.5 agents scan stack, scripts and risky paths in parallel
▸ every agent returns a summary, never raw files
▸ Opus 5.5 writes CLAUDE.md, the agents, permissions, the hook and the skill
▸ anything touching permissions comes back to you before it is written
▸ JEV resolves micro-forks in code and escalates only the split ones

paste the tree and this prompt into Claude Code ↓

"Set up this repo for me:
1. First scan it with 3 Haiku 5.5 agents in parallel: one reads the stack and manifests, one reads the build and npm scripts, one finds paths that are dangerous to touch.
2. Each agent returns a summary only, never raw file contents.
3. Then write CLAUDE.md, 3 subagents, a permissions file, a format hook and a /ship skill, all based on what the scan actually found, not on defaults.
4. Every risky path from the scan goes into the deny list before anything runs.
5. Ask me before writing anything that grants permissions.
Show me every file as a diff first. No writes until I say go."

↳
https://t.co/GoHZ0potb7

save it, this is the first thing you run in every new repo.

[Akshay 🚀 @akshay_pachaar]

4 types of LLM caching, clearly explained.

(KV, prefix, prompt, and semantic caching)

An agent can send the same system instructions, tool definitions, and conversation history on every turn. Without reuse, the model processes that shared context again.

That initial prompt processing is called prefill. It contributes to the wait before the first token appears.

The four caches in the graphic reduce repeated work at different points in this process.

𝟭) 𝗞𝗩 𝗰𝗮𝗰𝗵𝗶𝗻𝗴 𝘀𝗮𝘃𝗲𝘀 𝘄𝗼𝗿𝗸 𝗯𝗲𝘁𝘄𝗲𝗲𝗻 𝗴𝗲𝗻𝗲𝗿𝗮𝘁𝗲𝗱 𝘁𝗼𝗸𝗲𝗻𝘀.

During generation, the model stores keys and values, the vectors attention uses to retrieve information from earlier tokens. Each new token reuses those vectors and adds its own, avoiding repeated computation at the cost of memory.

𝟮) 𝗣𝗿𝗲𝗳𝗶𝘅 𝗰𝗮𝗰𝗵𝗶𝗻𝗴 𝘀𝗮𝘃𝗲𝘀 𝘄𝗼𝗿𝗸 𝗮𝗰𝗿𝗼𝘀𝘀 𝗿𝗲𝗾𝘂𝗲𝘀𝘁𝘀.

When a request begins with an already processed sequence, the server can reuse its saved attention state and process the remaining suffix. This helps when requests share instructions, documents, or conversation history.

In vLLM, each cached block’s identifier incorporates the preceding prefix, so identical text after a changed prefix is insufficient for reuse.

𝟯) 𝗣𝗿𝗼𝗺𝗽𝘁 𝗰𝗮𝗰𝗵𝗶𝗻𝗴 𝗽𝘂𝘁𝘀 𝘁𝗵𝗲 𝗽𝗿𝗼𝘃𝗶𝗱𝗲𝗿 𝗶𝗻 𝗰𝗵𝗮𝗿𝗴𝗲.

A provider manages this prefix reuse through its API, including cache lifetimes and billing rules. Eligible repeated input receives discounted pricing, while the model still generates a fresh answer.

Put stable instructions and documents at the beginning, followed by the changing question, to preserve that reusable prefix.

𝟰) 𝗦𝗲𝗺𝗮𝗻𝘁𝗶𝗰 𝗰𝗮𝗰𝗵𝗶𝗻𝗴 𝗿𝗲𝘂𝘀𝗲𝘀 𝘁𝗵𝗲 𝗮𝗻𝘀𝘄𝗲𝗿 𝗶𝘁𝘀𝗲𝗹𝗳.

The system converts a prompt into an embedding, a numerical representation of its meaning, and searches for similar stored prompts. An accepted match returns the saved answer without calling the generation model, saving output tokens too.

This suits repeated questions with stable answers, but similarity alone cannot establish that an answer applies to the current user or situation.

These techniques can work together. A semantic cache miss can still benefit from a cached prompt prefix, then use KV caching during generation.

Prefix caching and provider prompt caching share the same underlying reuse. Semantic caching adds a different decision, whether a previous answer is still valid for this request.

That said, I also wrote a full breakdown of KV caching. The article is quoted below.

Hope you enjoyed reading.

Cheers! :)

[DeepLearning.AI @DeepLearningAI]

A new agentic research model and harness called AREX checks its answers requirement by requirement, keeps what's verified, and researches only the gaps. 📰✅

📊 82.5% BrowseComp, 82.0 F1 WideSearch-en
📈 Harness alone: up to +10 pts
🧠 Fine tuned 4B beat untuned 35B on 5 of 6 benchmarks

https://t.co/DsoUUOaLH5

#DeepLearningAI #AIAgents #LLMs

[Jay Alammar @JayAlammar]

Figure 1-20 is the most important figure out of the 330+ in An Illustrated Guide to AI Agents.

1- It shows what today's LLM-based AI agents are composed of, including their user and environment loops.

2- It is also therefore a table of contents for Part 1: Agent Fundamentals, showing where each of the key components is explained in details.

Read it here:
Amazon: https://t.co/0dX0Js2IP8
O'Reilly: https://t.co/rrf0n2l3QG
PDF e-book: https://t.co/AThVjkVkB4

[Arindam Majumder 𝕏 @Arindam_1729]

Someone gave Jev a crawler + one GPT agent and asked:

“Where is the online money right now?”

It read 5,137 open jobs and found 3 niches where clients are paying and almost nobody is bidding.

Everyone asks ChatGPT for side hustle ideas.

Everyone gets the same 10 answers.

This approach is different.

The crawler reads what clients are actually paying for. Jev analyzes thousands of jobs cheaply and GPT only gets called when something looks worth researching.

The result:

01. Chrome extensions → $545 median, 11 bids/job
02. TikTok Shop setup → $232 median, 6.5 bids/job
03. Notion setup → $181 median, 3 bids/job

Jev made 9,617 calls for $0.13.

GPT alone would have cost ~$58.66.

The whole run cost $6.53.

The interesting part isn’t even the 3 niches.

It’s the architecture:

Crawler → Jev filters → Jev ranks → GPT researches → real opportunities

Cheap model does the volume. Expensive model only handles the work that needs deeper reasoning.

That’s a much more interesting way to use AI for finding opportunities.

[kiosa @thegreatest_sv]

whoever built this realized we’ve been using our smartest AI for the dumbest possible jobs

your $20-$200/mo frontier model is researching, writing code, planning…

and then you’re paying the same brain to decide YES / NO.

Jev flips that.

it’s a tiny decision model that sits in front of the expensive stuff and handles the boring forks:

-> which agent goes next?
-> is this worth researching?
-> reply or ignore?
-> publish, wait or escalate?
-> BUY / SELL / HOLD?
-> click this or keep looking?

then I saw the numbers people are reporting and the whole architecture started making a lot more sense.

~1,000 papers classified for around $0.08
~500 emails sorted for around $0.035

browser-agent loops demonstrated around 7 seconds for ~$0.0039

and at the quoted pricing, roughly 10,000 ~1K-token decisions comes out around $0.42.

that’s not because Jev became smarter than Claude.

it’s because nobody asked it to be Claude.

the setup is stupidly simple:

-> big LLM gets the hard thinking, research and writing
-> Jev gets the thousands of tiny decisions between those steps
-> code actually clicks, saves, sends and executes

and suddenly you start noticing these little decisions everywhere.

> your inbox.
> X feed.
> support queue.
> meetings.
> leads.
> video clips.
> browser agents.
> even paper-trading loops.

that’s the part I think most agent builders are missing.

the expensive model doesn’t need to touch every step just because it’s capable of touching every step.

> let Claude think.
> let Jev decide what deserves Claude.
> let code do the actual work.

I broke down 12 places you can wire this into an agent today + the exact router setup in the article below

[clem 🤗 @ClementDelangue]

We turned Claude Code, Codex, Hermes, Pi, @opencode and other coding harnesses into RL environments. No changes to the harnesses, no changes to the training code. Any open model, any task set, fully open source my friends!

Same model, same weights: 62% under Mini-SWE-Agent, 33% under Claude Code. But training inside a real harness normally means reimplementing it as an environment, so most models get trained in a scaffold nobody actually ships.

The fix is a proxy, not a rewrite. The harness thinks it's talking to a model API. It's actually talking to a capture proxy that speaks the 4 formats coding agents use (OpenAI Chat Completions, OpenAI Responses, Anthropic Messages, Gemini), forwards to @vllm_project, records the exact token IDs and logprobs vLLM sampled, and hands TRL sequences it can train on. The harness becomes the environment. 10 harnesses run through it today, none modified.

And because you control the reward, you can shape behavior the harness never asked for. We added a small bonus for solving a task in fewer tool calls: on tasks it already solved, the model now uses 31% fewer calls, in every harness, and about half under Codex.

Tested on LFM2.5-2.6B from @liquidai:
→ Train in one harness: better mostly in that harness (OpenCode 34% → 58%).
→ Train in 4 at once: better in all 4 (42% → 54%).
→ SFT on 3,189 rollouts from Qwen3.8-27B instead: plateaus at 47.5%, below both RL runs.

Everything is open and reproducible: the capture proxy in OpenEnv, the trainer in TRL, the tasks, the SFT data, the training code and all 7 trained models. Bigger models and bigger runs next.

Full guide: https://t.co/sKZURuOcza

[Defileo🔮 @defileo]

https://t.co/N0MN5hBgHT

[Kirk Borne @KirkDBorne]

Handbook on Understanding Harness Engineering for AI Agents [48-page PDF]: https://t.co/mHjkhT405C

[Dilum Sanjaya @DilumSanjaya]

Industrial software doesn’t have to feel like industrial software

Manage warehouses like you’re playing a strategy game

Built with Opus 5.5

[ℏεsam @Hesamation]

found this free Harness Engineering course on GitHub with 14 lectures + 8 hands-on projects. it covers:

> how Claude Code, Codex, DeepSeek build their harness
> loop & graph engineering
> automated agent loops
> observability, testing and verification
really good source for learning harness essentials by doing.

https://t.co/ebOTeIRKuy

[Andrej Karpathy @karpathy]

We'll be spending a lot more time trying to understand the outputs of language models. A few thoughts, tips & tricks:

Writing. Something I've had success with: Ask your LLM to explain something in ASD-STE100, it's a controlled language specification originally developed for aerospace maintenance documentation. LLMs well-versed in this language and it comes with heavy constraints on clean writing style that I often find a lot more readable. Sometimes I've tried to soften it a bit e.g. ask for "80% of the way to ASD-STE100" because the spec is quite stringent. But even better:

Diagrams / images. Instead of writing, ask your LLM to create a diagram. These can be a lot easier to process, parse, and understand. But even better:

Web pages. Ask for output "in HTML" to get a beautiful, interactive webpage. LLMs are getting really good at frontend and can create beautiful experiences, animations, etc. But even better:

Explainer videos. The output format I am most bullish on is fully custom / bespoke explainer videos generated on any arbitrary topic. Experiment with things like "Create a 3b1b style video explainer on X. Use my ElevenLabs API key for audio narration". (you'd need an API key for the latter or you can ask your LLM to find you decent free alternatives that use your local compute). This is actually starting to work!

In summary:
- As LLMs get better, they will do more and more of the legwork autonomously, and a lot more of our work will rise up the abstractions into oversight and understanding.
- Luckily, LLMs can help here too because as intelligence and code are increasingly abundant, you can ask for large, custom, discardable software artifacts (e.g. web apps, video explainers) that would have never made sense to create before. Push the boundaries here and you'll be surprised.

[Colin Jarvis @colintjarvis]

What does it take to be a great FDE at OpenAI - we get this question alot, so I sat down with @gregorojstersek to talk about how we’ve built the team, what the work looks at, and the values we want the team to have - always make the thing exist (and work) first, then look for scale second.

We talk through a few of the recent focuses in how we operate - we've got a bigger focus on domain experts than we used to, and ultimately success for us revolves around working to replace ourselves in every customer problem we take on.

Check out the interview below, and if you're interested in the role then I've included the careers page too - look forward to seeing what you build!

Interview: https://t.co/U4ifNPHTIB
FDE careers: https://t.co/NmuUTqaHvs

[Jackson Hinkle 🇺🇸 @jacksonhinkle]

NEVER FORGET this HERO!