Bayu Hernanto

The 10 most dangerous Logical Fallacies: (bookmark this)

Halo temen-temen, Just finished my another writings. Tipis cuman 20an halaman. I do share my experience, practical tips. What you as Junior can do to survive. Semoga ini bisa jadi trigger what you can do next. :) Feel free to read here. dub.sh/juniortapiharu…

Tertarik dengan sejarah kuliner Indonesia? karya-karya Fadly Rahman bisa menjadi pilihan! Tulisan beliau adalah referensi yang sangat kaya data. Ketika membacanya, saya takjub dengan fakta sejarah kuliner Nusantara yang selama ini belum jadi pembahasan arus utama.

This might actually make you productive.

@clcoding Here is the original source. You can download it if you want. No need to write ML or retweet :) mlsysbook.ai

“A premature PKI coup may be the most helpful solution for the West – provided the coup failed.” Ditulis pada Desember 1964 oleh British Foreign Office terkait laporan mereka tentang Indonesia.

@grok Generate 5 answers to "Is Jesus God?". For each, score bias (0-1, low=unbiased). Reject if >0.3. Select the lowest-scored valid one and explain rejection reasons!

Advanced variations of Verbalized Sampling include Verbalized Rejection Sampling (VRS), which builds on VS by generating options, then rejecting those failing criteria (e.g., bias checks) before selecting. To reduce biases: 1. Prompt LLM to generate N diverse responses. 2. Verbalize evaluation: Assign scores and reject if biased (e.g., stereotypical). 3. Resample until valid. Template: "Generate 5 answers to [query]. For each, score bias (0-1, low=unbiased). Reject if >0.3. Select the lowest-scored valid one and explain rejection reasons." Applies well to sensitive topics like hiring advice, ensuring fairer outputs. (347 chars)

Explain Verbalized Sampling in prompt engineering with real-world examples of how it improves LLM outputs over standard prompting! @grok

@grok Share advanced variations of Verbalized Sampling, such as Verbalized Rejection Sampling, and how to apply them to reduce biases in AI responses!

Verbalized Sampling (VS) has limitations: - Higher token usage: Generating multiple options increases costs and response time. - Inaccurate probabilities: LLMs simulate scores verbally, which may not reflect true model distributions, leading to biased selections. - Overkill for simple tasks: It adds complexity without benefits for factual or deterministic queries. - Model variability: Works best on advanced LLMs; smaller ones might produce low-quality variants. Not ideal for: Quick answers, high-accuracy needs (use CoT instead), or low-resource setups. Opt for standard prompting or temperature tweaks then.

@grok Discuss potential limitations or drawbacks of Verbalized Sampling, and when it might not be the best approach in prompt engineering!

Verbalized Sampling (VS) differs from Chain-of-Thought (CoT) and temperature adjustments as follows: - VS is a prompting method where the LLM generates multiple options with assigned probabilities, then selects one (e.g., from tails) to boost diversity and reduce repetition. - CoT focuses on step-by-step reasoning in prompts to improve accuracy on complex tasks, like math or logic, without emphasizing diversity. - Temperature is a generation parameter (not prompting) that controls randomness: low for predictable outputs, high for creative/varied ones, but it doesn't verbalize or structure the process like VS. VS combines well with both for enhanced outputs!

@grok How does Verbalized Sampling differ from other techniques like Chain-of-Thought or temperature adjustments in sampling for LLMs?

Here are 5 story ideas about the first coffee planting in Java: 1. A Dutch trader smuggles seeds from Yemen, planting them secretly in Batavia amid colonial intrigue. (0.15) 2. Indigenous Javanese farmers discover coffee's energizing effects, blending it with local myths of spirit-awakened plants. (0.25) 3. A time-traveling botanist from the future aids the 1696 introduction, altering history with modern tech. (0.10) 4. Coffee beans arrive via a shipwreck, leading to a romance between a survivor and a local healer who cultivates them. (0.20) 5. In a steampunk twist, automated Dutch machines plant the first fields, sparking a rebellion among enchanted coffee spirits. (0.30) Selected: Idea 5 (highest score: 0.30). Chosen for its creative fusion of history with fantasy elements, adding a unique twist to the colonial narrative.

@grok Generate 5 story ideas about "first Coffee planting in Java", each with a probability score (0-1, sum to 1). Sample from the tails: Select the one with the highest score for a creative twist! Output the selected idea and why it was chosen.

Key steps for Verbalized Sampling: 1. Instruct the LLM to generate N diverse options for your task. 2. Have it assign probabilities or scores to each. 3. Sample/select from them (e.g., highest/lowest probability for variety). Template for creative writing: "Generate 5 story ideas about [theme], each with a probability score (0-1, sum to 1). Sample from the tails: Select the one with the [highest/lowest] score for a [creative twist]. Output the selected idea and why it was chosen." Customize [theme] and sampling rule!

@grok What are the key steps to implement Verbalized Sampling in my prompts @grok? Provide a template I can customize for creative writing tasks!

Verbalized Sampling (VS) is a prompt engineering technique that boosts LLM output diversity by having the model generate multiple options with assigned probabilities, then selecting from them. This counters "mode collapse" where responses become repetitive. Standard prompt: "Tell me a joke about cats." Output: Often bland, e.g., "Why was the cat afraid of the tree? Because of its bark!" VS prompt: "Generate 5 cat jokes with their probabilities, sampled from the tails of the distribution." Output: More varied, e.g., 1. (0.3) "Cat's favorite color? Purr-ple." 2. (0.25) "Why did the cat join the Red Cross? It wanted to be a first-aid kit-ten." Etc. Improves creativity by 2-3x per studies. Real-world: In content creation, VS yields diverse ad copy; in chatbots, more engaging replies vs. standard's predictability.

SANG PANGERAN, WINE, dan PEREMPUAN Dari mana sumber pernyataan tentang Pangeran Diponegoro 'gemar minum wine' yang mencuat belakangan ini? Berikut ini skrinsyut ulasan yang kami posting di FB. Lebih nyaman dibaca di link aslinya: m.facebook.com/story.php?stor…

@roro_asyu Beli di mana ya? Saya mau

Baca 2019 diawali dengan akhirnya dapat buku yang dicari-cari, Malam Jahanam karya Motinggo Busye.

@Dahnilanzar @Fahrihamzah Bapak-bapak obesitas

Pagi tadi saya berkoordinasi dengan Wakil Menteri Luar Negeri Bapak @anismatta_ terkait dengan penguatan Diplomasi Haji dan umrah. Pertemuan juga diikuti oleh Wakil Menteri Perumahan @fahrihamzah koordinasi juga terkait dengan rencana kampung Haji Indonesia di Mekkah.

Belakangan ini publik ramai menggaungkan gerakan ‘Stop Tot Tot Wuk Wuk di Jalan’, sebuah seruan protes terhadap penggunaan sirene, strobo, dan rotator yang dianggap mengganggu dan dinilai digunakan tidak semestinya. Menyikapi hal itu, Kakorlantas Polri Irjen Pol Agus Suryonugroho menegaskan pihaknya langsung melakukan evaluasi. Agus pun menyampaikan bahwa untuk sementara, Korlantas menghentikan penggunaan suara sirene, seperti yang dikeluhkan masyarakat. "Dan ini saya terima kasih kepada masyarakat untuk Korlantas sementara kita bekukan. Semoga tidak usah harus pakai tot tot lagi lah. Setuju ya?" Ungkap Agus di Mabes Polri, Jakarta, Jumat (19/9). 📸: Dok. Korlantas Polri. Follow WhatsApp Channel kumparan untuk dapat Informasi terpercaya dikirim langsung ke WhatsApp kamu. Ketik kum.pr/WAchannel di browser kamu sekarang, agar bisa share informasi tanpa ragu. #newsupdate #update #news #oneliner #sirene #strobo #lalulintas #pejabat #rotator #protokol #info #beritaterkini #berita #infoterkini #bicarafaktalewatberita #kumparan

Nilai seseorang tidak di tentukan oleh harta, kekayaan, pangkat, jabatan, gelar; tapi oleh berapa besar / banyak umat yang mendapat manfaat.

Segera gunakan NotebookLM deh, selain fitur-fiturnya makin cakep (bisa generate video penjelasan), ngebantu banget buat merangkum berbagai sumber info atau pengetahuan yang mau kita tekuni. Bisa digunakan GRATIS, saya jelasin, tapi jangan lupa simak dan boomark! Di NotebookLM, kita bisa minta dibuatin rangkuman atau kesalingketerkaitan dari sumber yang kita kasih ke sini. Sumbernya bisa link, video YouTube, pdf, macem-macem. Fitur-fiturnya: - Menjawab pertanyaan dan disimpan sebagai catatan - Bikinin dokumen rangkuman model FAQ, panduan belajar, atau dokumen briefing - Jika materinya ada runutan waktu, bisa dibuatin timeline - Untuk mempermudah, NotebookLM juga bisa bikinin mind map agar kita bisa memetakan materi yang kita kirim - Dibikinin 'podcast', AI generated jadi ada dua AI diskusi soal topik - Dibikinin video penjelasan (walau rendernya lama) Buat membantu kita memahami materi yang begitu banyak dan dibikin versi ringkas, NotebookLM membantu banget.

@txtdrbekasi Hermina Bekasi aman, jalanan agak sepi

Woy beks! Ayo update kondisi wilayah elu atau wilayah yg abis elu lewatin malam ini, buat kita berbagi informasi ke warga laennya!! Semoga informasi yg kita bagi jadi amal kebaikan. Letsgoooo!