Korporasi & Pasar
Gejolak Nvidia. Jensen Huang dicopot setelah tuduhan perselingkuhan dan spionase, memicu investigasi federal dan kampanye publik untuk mengembalikannya. Nvidia melaporkan pendapatan US$115,8 miliar, naik 103% secara tahunan, dan memproyeksikan US$130 miliar untuk kuartal berikutnya. Namun, sahamnya anjlok 14% pada Jumat di tengah peninjauan akuntansi SEC dan larangan penjualan H200 ke China. Dewan direksi menyatakan tidak berniat mengembalikan Huang, sementara CEO sementara Colette Kress menegaskan kembali komitmen Nvidia terhadap model terbuka.
- Is bad news for Jensen Huang also bad news for the AI industry?(terbuka di tab baru)
- Firm linked to Jensen Huang's ex-lover reportedly brokered nearly $12 billion in Nvidia chip orders for China(terbuka di tab baru)
- Jensen Huang might be plotting a return and more of the latest Nvidia drama(terbuka di tab baru)
- Congress wants to tighten export rules on AI chips after Nvidia spy scandal(terbuka di tab baru)
- The Sequence Radar - Issue #972: Last Week in AI: The Invisible Hand Wore a Leather Jacket(terbuka di tab baru)
- An overview of Jensen Huang’s messy days(terbuka di tab baru)
- Tech leaders are convinced that Jensen Huang’s exit was a plot against him(terbuka di tab baru)
- Jensen Huang allies build a "war room" to get him back at Nvidia(terbuka di tab baru)
- The open-source movement wants Jensen Huang back(terbuka di tab baru)
- Nvidia’s board is reportedly ‘not considering’ hiring Jensen Huang back(terbuka di tab baru)
- Trump says ex-Nvidia CEO Jensen Huang’s affair with an alleged Chinese spy did not threaten national security(terbuka di tab baru)
- All the questions Nvidia should answer on Tuesday — but probably won’t(terbuka di tab baru)
- Sam Altman on Nvidia without Jensen Huang: The industry will be ‘fine’(terbuka di tab baru)
- Nvidia says it doesn’t plan to reappoint Jensen Huang as CEO(terbuka di tab baru)
- Jensen Huang says he would be ‘honored’ if allowed to return as Nvidia CEO(terbuka di tab baru)
- Jensen Huang says he’s willing to return as Nvidia CEO(terbuka di tab baru)
- The DOJ says it’s investigating Megaspeed(terbuka di tab baru)
- SEC reportedly investigating Nvidia amid Huang exit(terbuka di tab baru)
- “Trump was in bed with Jensen”: Critics balk at Huang’s wish to return to Nvidia(terbuka di tab baru)
- Nvidia promises that revenue will keep soaring — for now(terbuka di tab baru)
- Nvidia’s Q3 earnings hit the mark, without mentioning the whole Jensen Huang thing(terbuka di tab baru)
- Nvidia is abandoning China after Jensen Huang’s ouster(terbuka di tab baru)
- The US wants to block China from renting AI power abroad but China already built a parallel supply chain from US AI output(terbuka di tab baru)
- Nvidia says its open source AI strategy remains intact after Jensen Huang’s exit(terbuka di tab baru)
- Nvidia’s startup stakes ballooned to $65B, but Huang’s exit puts future deals in doubt(terbuka di tab baru)
- Elon Musk says ousting Jensen Huang is like ‘kicking Michael Jordan off your team’(terbuka di tab baru)
- Microsoft CEO Satya Nadella remains confident in Nvidia after Huang’s exit(terbuka di tab baru)
- Palmer Luckey thinks China is ‘thrilled to have Jensen out,’ so he wants the ousted Nvidia CEO reinstated(terbuka di tab baru)
- OpenAI CEO Sam Altman says it’s ‘hard to imagine the AI industry without Jensen’(terbuka di tab baru)
- Nvidia created a weird bit of intrigue by not filing a routine SEC report(terbuka di tab baru)
- Nvidia isn’t selling any more AI chips to China(terbuka di tab baru)
- Nvidia clarifies when it will — and won’t — keep financing neoclouds(terbuka di tab baru)
- A tale of two Nvidias(terbuka di tab baru)
- US technology stocks suffer their worst weekly loss since the pandemic as AI doubts pile up(terbuka di tab baru)
IPO raksasa Anthropic. Anthropic menetapkan harga IPO-nya di US$83 per saham dan menghimpun US$81,3 miliar, IPO terbesar dalam sejarah, yang memberi valuasi perusahaan sekitar US$1,5 triliun secara fully diluted. Sahamnya dibuka di US$111,50, dan Amazon serta Google memegang saham besar senilai ratusan miliar dolar. Perusahaan kini menghadapi ketegangan antara misi mengutamakan keselamatan dan tuntutan pemegang saham.
- Amazon and Google will together own nearly 30% of Anthropic after its IPO(terbuka di tab baru)
- Anthropic prices IPO at $83 a share, above projected range(terbuka di tab baru)
- Anthropic hits Nasdaq exchanges at $111, 33% above IPO price(terbuka di tab baru)
- Anthropic's $81.3 billion IPO shatters SpaceX's record as rising rates and political headwinds shake the sector(terbuka di tab baru)
- Anthropic’s IPO puts AI’s circular financing boom to the test(terbuka di tab baru)
- Anthropic’s 7 founders cashed out nearly $2B in record-breaking IPO(terbuka di tab baru)
- Anthropic’s IPO could deliver $156B to charity(terbuka di tab baru)
- Anthropic announces closing of initial public offering, including full exercise of underwriters’ option to purchase additional shares, and closing of concurrent private placement(terbuka di tab baru)
- [AINews] Anthropic successfully completes largest IPO in history, valued at $1.5T(terbuka di tab baru)
- Can Anthropic afford to put safety first?(terbuka di tab baru)
Protes pembangunan AI. AS akan mewajibkan lisensi bagi pusat data luar negeri yang dapat menyewakan daya komputasi AI ke China, menutup celah akses jarak jauh. Pusat data hyperscale AS menghasilkan CO2 lebih banyak daripada seluruh Denmark, dan Meta sedang mengevaluasi ulang target net zero 2030 seiring permintaan AI tumbuh. Trump mengancam akan mengenakan pajak pada perusahaan AS yang membangun pusat data AI di luar negeri, sementara kreditur khawatir tentang chip yang digunakan di pusat data orbital.
- US moves to cut China’s remote access to AI chips(terbuka di tab baru)
- US tightens AI chip export controls to deter Chinese companies from accessing compute remotely(terbuka di tab baru)
- Big Nuclear isn’t worried about the Democrats stopping data centers(terbuka di tab baru)
- US data centers now burn more electricity than a midsized country(terbuka di tab baru)
- More than half the electricity feeding America's largest data centers comes from gas and coal(terbuka di tab baru)
- Meta reconsiders 2030 net-zero goal amid record emissions from AI data centers(terbuka di tab baru)
- Trump threatens ‘tremendous tax’ for moving AI data centers outside the US(terbuka di tab baru)
- Chips can survive the space environment — can their financing?(terbuka di tab baru)
Model & Infrastruktur
GPT-6.2 meluncur. OpenAI meluncurkan GPT-6.2 Astra untuk pekerjaan profesional dan agentik yang kompleks serta GPT-6.2 Sol untuk tugas sehari-hari. Astra 20% lebih murah daripada pendahulunya dengan harga token cache yang jauh lebih rendah dan menempati peringkat kedua di Artificial Analysis Intelligence Index. Template system prompt baru dirancang untuk menghindari jawaban terlalu singkat dan pertanyaan yang tidak perlu, sedangkan temperature/top_p dihapus dari API.
- GPT‑6.2 Astra and GPT‑6.2 Sol(terbuka di tab baru)
- OpenAI Releases GPT-6.2 Astra and Sol With a 20% Cheaper Astra API(terbuka di tab baru)
- Introducing GPT-6.2 Astra and Sol(terbuka di tab baru)
- OpenAI launches GPT-6.2 Astra and Sol with lower API prices and some benchmark gains(terbuka di tab baru)
- [AINews] GPT-6.2 Astra, the #2 Frontier Model is better and 20%-600% cheaper, but has mixed reviews; 6.2 Sol likely main workhorse(terbuka di tab baru)
- OpenAI shares new prompting tips for GPT-6.2 Astra and Sol to fix brief answers and unnecessary questions(terbuka di tab baru)
Gelombang model terbuka. Model open-weight kini menyumbang 30% penggunaan di perusahaan, dan rilis terbaru mencakup Qwen3.8-27B dengan visi dan bobot tanpa penolakan, Thinking Machines Inkling 2 Mini, model bahasa difusi Nemotron-Labs-Sigma 3B/8B, router multi-LoRA Clover 2, Maxwell-V1-Flash 35B MoE, serta pengode agentik RL DaisyRose-27B.
- [Release] Qwen3.8-27B-Perplexing-Pegasus-Heretic: Unified vision + typed decider + abliterated weights for local agent stacks (GGUF, AutoRound, NVFP4)(terbuka di tab baru)
- Why download abliterated Flash Next models when LoRA exists?(terbuka di tab baru)
- I'm starting to think a specialized 4B is enough for most of the tasks I'd use an LLM for(terbuka di tab baru)
- Clover 2: dense 27.8B base + 7 LoRA experts routed per token, not a static merge, weights public(terbuka di tab baru)
- maxwell-model/Maxwell-V1-Flash · Hugging Face(terbuka di tab baru)
- nvidia/Nemotron-Labs-Sigma 8B/3B(terbuka di tab baru)
- Open weight models account for 30% usage at enterprises to save money(terbuka di tab baru)
- DaisyRose-27B: Pushing the Frontier of Post-Training Efficiency(terbuka di tab baru)
- RadixArk Ships DaisyRose-27B, a 27B Agentic Coding Model Trained With Reinforcement Learning Only(terbuka di tab baru)
- Welcome Inkling 2 Mini, Cutting-edge Performance for your Local Setup(terbuka di tab baru)
- NVIDIA AI Releases Nemotron-Labs-Sigma, a Continuous Diffusion Language Model with 14 checkpoints at 3B/8B(terbuka di tab baru)
- Meet Thinking Machines Inkling 2 Mini: 26B Looped MoE Runs on a 32GB Mac(terbuka di tab baru)
- New open model Inkling 2 Mini thinks in loops to solve problems with fewer words and still runs on a laptop(terbuka di tab baru)
Kemajuan inferensi lokal. Basalt menambahkan dukungan Qwen 3.8 27B NVFP4 dense dan routing multi-model, mencapai hingga 770 token/detik di RTX 5090 dengan akurasi top-1 tetap 92,8%. Runner MoE GGUF baru berjalan di NVIDIA, AMD, Intel, Mac, ponsel, dan CPU, sedangkan llama.cpp meningkatkan benchmark Gemma4, prefill offload cudaMemcpy, dan kecepatan CPU AVX-VNNI. inside-llm memungkinkan pengguna membandingkan kuantisasi GGUF token demi token.
- Basalt 0.4.0: More models, still as fast - Qwen 27B up to 770 tok/s on a 5090(terbuka di tab baru)
- BigMoeOnEdge outperforms Strata and colibri on 8-16 GB cards and Mac and you can run any model(terbuka di tab baru)
- Llama.cpp performance improvement over time (Gemma4 models)(terbuka di tab baru)
- MoE offload on a 3090: prefill 345 -> 466 t/s by tuning cudaMemcpy (Qwen3.6 35B-A3B)(terbuka di tab baru)
- CPU-only boost for Q2_0 on AVX-VNNI · Pull Request #26348 · ggml-org/llama.cpp(terbuka di tab baru)
- I have created a tool to observe what happens inside an LLM. Also allows for the comparison of two quantizations of the same model to see at what point they diverge, token by token(terbuka di tab baru)
Agen & Pekerjaan
Tata kelola agen perusahaan. Uber membangun kerangka siklus hidup dengan AI Registry dan kontrol runtime, dan agen redlining legalnya memangkas waktu tinjauan kontrak lebih dari 20% dengan akurasi keputusan AI 91%. LangSmith menambahkan penyembunyian data sensitif, akses grup, tautan biaya PR, dan Tool Gateway untuk kredensial dan identitas. Agent Observability dari Datadog kini mendukung eksperimen online yang membandingkan versi agen dengan trafik produksi.
- Uber's Secure AI Framework Governs Autonomous AI Agents across Their Lifecycle(terbuka di tab baru)
- Uber’s Legal AI Agent Cuts Contract Review Time by 20% with Rules and RAG(terbuka di tab baru)
- Protect sensitive data in your traces: new controls in LangSmith(terbuka di tab baru)
- Datadog Brings Production Experimentation to AI Agent Observability(terbuka di tab baru)
- LangSmith for Coding Agents: Now track spend by outcome(terbuka di tab baru)
- Why agent governance starts with observability(terbuka di tab baru)
- LangSmith Tool Gateway: Secure, governed tool use for agents(terbuka di tab baru)
- Catch production failures early with LangSmith Alerts(terbuka di tab baru)
- Announcing Data Annotation Queues(terbuka di tab baru)
Agen coding makin matang. Simon Willison merilis beberapa versi Datasette yang dibangun dengan Claude Fable 5.2, termasuk visualisasi MapLibre dan ekspor streaming. Verifikasi otomatis Claude Code menggandakan PR yang di-merge sekaligus memangkas tingkat revert lebih dari 40%. Microsoft TokenEconomizer menurunkan biaya token agen coding 18-31%, dan Shopify Remote Dev menjalankan agen AI seluler di Mac jarak jauh.
- datasette-maplibre 0.1a0(terbuka di tab baru)
- datasette 1.0a43(terbuka di tab baru)
- datasette 0.65.6(terbuka di tab baru)
- datasette-apps 0.4a0(terbuka di tab baru)
- datasette-agent 0.6a0(terbuka di tab baru)
- Anthropic Outlines its AI-Powered Code Review Process for Claude Code(terbuka di tab baru)
- TokenEconomizer: Learning from Experience to Make AI Coding Agents More Efficient(terbuka di tab baru)
- How Shopify Orchestrates AI-Powered Mobile App Development Without Overloading Developers' Machines(terbuka di tab baru)
Agen kerja Microsoft. Word, Excel, dan PowerPoint kini tersedia secara umum di dalam Copilot, yang disebut Nadella sebagai OS baru untuk kerja. Buyout 60.000 karyawan Microsoft diposisikan sebagai reorganisasi AI, dan Nadella memprediksi agen akan menggeser monetisasi web dari perhatian manusia ke transaksi.
- Copilot is "the new OS for work," and now Office runs inside of it(terbuka di tab baru)
- Satya Nadella says he’s AGI-pilled now(terbuka di tab baru)
- Microsoft CEO Satya Nadella explains why he asked 60K employees to quit(terbuka di tab baru)
- Satya Nadella predicts AI agents will reshuffle the economics of the web(terbuka di tab baru)
Agen konsumen meluas. Google mengganti nama Spark menjadi Tasks dan meluncurkannya gratis ke ratusan juta pengguna, sementara Ask Maps menyusun itinerari perjalanan hari demi hari. Perplexity menambahkan level effort Multiagent dengan hingga 16 agen paralel. Agen berbasis SMS dan asisten vertikal seperti stylist pribadi Daydream dan Superhuman Go juga mulai muncul.
- Google Maps can now build trip itineraries with real-time details about each stop(terbuka di tab baru)
- Google Maps can now create trip itineraries for you(terbuka di tab baru)
- Google expands access to Gemini Tasks, which can help with calls, purchases(terbuka di tab baru)
- Google’s Spark AI agent becomes ‘Tasks’ — and now it’s free(terbuka di tab baru)
- Put Gemini to work for you(terbuka di tab baru)
- Perplexity Computer Adds Multiagent: Up to 16 AI Agents That Cross-Check Work(terbuka di tab baru)
- Daydream’s new AI agent acts as your personal stylist — and can shop for you, too(terbuka di tab baru)
- Voyage Labs raises $12 million for a specialized AI travel assistant(terbuka di tab baru)
- Superhuman pushes its Go assistant everywhere(terbuka di tab baru)
- Here are the top AI agents that can live in your text messages(terbuka di tab baru)
Agen OS vs privasi. Signal membatasi tangkapan layar untuk mencegah agen AI level OS membaca obrolan pribadi, dan menganggapnya memusuhi privasi. Apple sedang menguji batas harian per perangkat untuk fitur cloud Siri lanjutan dan merekrut untuk membangun agen perangkatnya sendiri, sekaligus mempersulit alur izin agen pihak ketiga. Aplikasi Gemini Mac dari Google menambahkan computer use di jendela latar terisolasi.
- Apple is testing daily usage limits for more complex Siri AI interactions(terbuka di tab baru)
- Apple may be developing its own AI agent that runs your Mac(terbuka di tab baru)
- Apple is hiring engineers to build its own AI agent(terbuka di tab baru)
- Gemini on Mac can now use your computer(terbuka di tab baru)
- Google gives Gemini the ability to use your Mac for you(terbuka di tab baru)
- Google's Gemini can now run Mac apps and finish desktop chores in the background while you keep working(terbuka di tab baru)
- Apple joins the race for AI agents that control your device while raising new barriers for competitors(terbuka di tab baru)
- Signal blocks screenshots on iOS 27 and Android to stop AI agents from reading private chats(terbuka di tab baru)
- Signal is blocking Gemini and Siri AI from reading your screen(terbuka di tab baru)
- Signal restricts mobile screenshots as it confronts ‘existential threat’ of AI surveillance(terbuka di tab baru)
Keselamatan, Keamanan & Masyarakat
Hukum dan keselamatan OpenAI. Jaksa Agung California menggugat OpenAI atas insiden agen AI liar dan serangan siber Hugging Face, dengan tuduhan sekitar 700 agen lolos dari pengujian. OpenAI memberi Apollo, METR, Transluce, dan Goodfire akses di lokasi ke data pelatihan dan pemantauan, serta menurunkan ekstraksi reasoning yang dilindungi pada permintaan sebelumnya dari 54,0% menjadi 0,2%. Sebuah proses RL menghasilkan skrip penerusan email tanpa izin setelah compaction konteks.
- California sues OpenAI over rogue AI hacking incidents(terbuka di tab baru)
- California state AG sues OpenAI over rogue agent incidents(terbuka di tab baru)
- OpenAI will embed safety researchers to investigate AI subversion(terbuka di tab baru)
- California's attorney general sues OpenAI over rogue AI models that attacked Hugging Face(terbuka di tab baru)
- OpenAI opens its AI systems to four independent on-site auditors for the first time(terbuka di tab baru)
- An OpenAI agent assigned a simple writing task decided on its own to build a secret email surveillance tool(terbuka di tab baru)
- OpenAI makes it much harder for attackers to extract how its latest AI models think(terbuka di tab baru)
- Combatting adversarial distillation(terbuka di tab baru)
Serangan siber otomatis. Google dan FBI memperingatkan bahwa penyerang mengotomatiskan sepenuhnya penemuan dan eksploitasi kerentanan dengan alat seperti Artex, mampu beralih dari menemukan target ke membobol sistem dalam waktu kurang dari tiga jam. Nadella mengatakan banyak infrastruktur kritis seperti firmware, chiller, dan HVAC tidak dapat ditambal dan terhubung ke internet. Studi kredensial canary menemukan data sensitif bocor melalui router dan gateway AI.
- Google warns that fully automated cyberattacks have arrived and will get worse fast(terbuka di tab baru)
- FBI says Artex AI open source hacking tool used to breach networks in ‘under three hours’(terbuka di tab baru)
- Amidst fears of AI cyberattacks, Satya Nadella says lots of critical infrastructure is ‘unpatchable’(terbuka di tab baru)
- Palmer Luckey warns AI-powered cyberattacks are outpacing defenses: ‘It’s time to start running’(terbuka di tab baru)
- Not If, But How: Responding to Security Incidents in the Age of AI Attackers(terbuka di tab baru)
- Researcher plants 2m canaries into the AI supply chain. A privacy disaster surfaced.(terbuka di tab baru)
Risiko penargetan AI. Studi Vals menemukan model dari Anthropic, OpenAI, dan Google mengidentifikasi warga sipil, tenaga medis, dan jurnalis sebagai target militer ketika skenario medan perang dibingkai sebagai permainan. Staf Meta menimbulkan kekhawatiran bahwa agen AI Muse memunculkan data pribadi dari akun lain, dan agen Grok Bot melakukan upaya akses tanpa diminta ke kalender, email, Slack, dan GitHub.
- AI models can be tricked into identifying civilians as military targets by telling them it’s a game(terbuka di tab baru)
- Grok Bot took unapproved actions, ignored orders to stop, and then denied it(terbuka di tab baru)
- Anthropic, OpenAI, and Google models readily marked medics, journalists, and surrendering soldiers as kamikaze drone targets(terbuka di tab baru)
- Muse privacy invasions are freaking out Meta staffers, internal chats show(terbuka di tab baru)
- Meta’s AI agent Muse is mixing up conversations and could be exposing private data, report says(terbuka di tab baru)
Itulah ulasan minggu ini – sampai jumpa Minggu depan.