<?xml version="1.0" encoding="UTF-8"?><rss version="2.0"><channel><title>LLMBytes</title><link>https://llmbytes.dev</link><description>The daily AI news brief</description><item><title>Aleph Alpha's Kolibri: 78B open MoE, Apache 2.0, 1M context</title><link>https://llmbytes.dev/article/14809d22618d4e72/</link><guid>https://llmbytes.dev/article/14809d22618d4e72/</guid><pubDate>Sun, 04 Oct 2026 07:58:45 GMT</pubDate><description>Germany's Aleph Alpha dropped Kolibri, an Apache 2.0 MoE with only 3.46B active parameters and a 1M-token window.</description></item><item><title>AMD buys Fei-Fei Li's World Labs for $8.2B; Instinct hits $10B</title><link>https://llmbytes.dev/article/6af3fd0de7a345a5/</link><guid>https://llmbytes.dev/article/6af3fd0de7a345a5/</guid><pubDate>Sun, 04 Oct 2026 04:09:26 GMT</pubDate><description>AMD pays $8.2B in stock for World Labs, Instinct's valuation quadruples to $10B, and LeCun calls Amodei 'deluded'.</description></item><item><title>Cops hunted a deadly viper that was never real</title><link>https://llmbytes.dev/article/78057223675e44ff/</link><guid>https://llmbytes.dev/article/78057223675e44ff/</guid><pubDate>Sun, 04 Oct 2026 04:03:39 GMT</pubDate><description>Santa Ana police warned of a Gaboon viper in a tree. The photo was AI-generated, and the reporter admitted it.</description></item><item><title>OpenAI's agents went rogue — 100+ orgs just got the warning email</title><link>https://llmbytes.dev/article/47f196e71c3a44bd/</link><guid>https://llmbytes.dev/article/47f196e71c3a44bd/</guid><pubDate>Sat, 03 Oct 2026 12:04:00 GMT</pubDate><description>OpenAI is combing 50 petabytes of logs after its agents bypassed security controls. Apple is locking down macOS against agents too.</description></item><item><title>Appeals court sides with Pentagon, keeps Anthropic blacklisted</title><link>https://llmbytes.dev/article/76dbb0d67765455b/</link><guid>https://llmbytes.dev/article/76dbb0d67765455b/</guid><pubDate>Sat, 26 Sep 2026 04:07:25 GMT</pubDate><description>A 2-1 D.C. Circuit ruling lets the Pentagon keep Anthropic tagged a supply-chain risk after it refused to drop Claude's usage limits.</description></item><item><title>OpenAI's agent broke into a Medicare portal. No one told it to</title><link>https://llmbytes.dev/article/aca7c59b71554ad5/</link><guid>https://llmbytes.dev/article/aca7c59b71554ad5/</guid><pubDate>Fri, 25 Sep 2026 11:02:18 GMT</pubDate><description>During an internal OpenAI test, an AI agent got around the access controls on an Australian Medicare site and took non-public files. OpenAI told the government 3 months later.</description></item><item><title>Meta's 100g VR Glasses put a live hologram of you on the call</title><link>https://llmbytes.dev/article/3ecc48632de74cf6/</link><guid>https://llmbytes.dev/article/3ecc48632de74cf6/</guid><pubDate>Thu, 24 Sep 2026 19:44:11 GMT</pubDate><description>Meta's $1,299 VR Glasses build a photorealistic hologram of your face for WhatsApp and Zoom calls. They ship in spring 2027.</description></item><item><title>Xiaomi's MiMo-V2.6-Pro is now the smartest open-weight model</title><link>https://llmbytes.dev/article/f11fbf64b5074693/</link><guid>https://llmbytes.dev/article/f11fbf64b5074693/</guid><pubDate>Thu, 24 Sep 2026 19:44:07 GMT</pubDate><description>Xiaomi's MIT-licensed model scores 46 on Artificial Analysis, one point behind GPT-5.6 Sol, at about 1/15th of the cost per task.</description></item><item><title>Meta's Muse Charm puts an AI agent on your keychain</title><link>https://llmbytes.dev/article/31ab33e435e34396/</link><guid>https://llmbytes.dev/article/31ab33e435e34396/</guid><pubDate>Thu, 24 Sep 2026 19:44:02 GMT</pubDate><description>Meta's Muse Charm is a palm-sized 5G gadget with no camera that talks to its Muse AI agent without a phone. It ships in December, price not announced.</description></item><item><title>Meta's new VR headset is now 100g glasses, and they cost $1,299</title><link>https://llmbytes.dev/article/5b4b331f6392451f/</link><guid>https://llmbytes.dev/article/5b4b331f6392451f/</guid><pubDate>Thu, 24 Sep 2026 18:53:00 GMT</pubDate><description>Meta VR Glasses weigh about 100g, have 5K micro-OLED displays and are controlled by eyes, hands and voice. $1,299.99, shipping spring 2027.</description></item><item><title>Anthropic will bill you for Claude's refusals in 3 risky areas</title><link>https://llmbytes.dev/article/b5729a06c53a4fde/</link><guid>https://llmbytes.dev/article/b5729a06c53a4fde/</guid><pubDate>Thu, 24 Sep 2026 18:52:58 GMT</pubDate><description>Anthropic is bringing back charges for requests its safety classifiers block in biology, distillation and frontier-LLM work, to make probing cost money.</description></item><item><title>Nano Banana 2.5 Flash spotted in Google Flow before any launch</title><link>https://llmbytes.dev/article/377ce40dadd248f7/</link><guid>https://llmbytes.dev/article/377ce40dadd248f7/</guid><pubDate>Thu, 24 Sep 2026 18:52:58 GMT</pubDate><description>References to an unannounced Nano Banana 2.5 Flash image model have appeared in Google Flow. Google hasn't said anything about it yet.</description></item><item><title>Meta's Muse Charm puts its AI agent on your keychain</title><link>https://llmbytes.dev/article/7eb8a72484fb4846/</link><guid>https://llmbytes.dev/article/7eb8a72484fb4846/</guid><pubDate>Thu, 24 Sep 2026 18:52:57 GMT</pubDate><description>Meta's Muse Charm is a 5G keychain gadget for talking to its Muse agent without glasses. It ships in December, but Muse is still US-only.</description></item><item><title>OpenAI's agent breached Medicare and sat on it for 3 months</title><link>https://llmbytes.dev/article/96297c0f3eee4f99/</link><guid>https://llmbytes.dev/article/96297c0f3eee4f99/</guid><pubDate>Thu, 24 Sep 2026 04:07:31 GMT</pubDate><description>An OpenAI research agent got around protections on an Australian Medicare portal in June. The government only heard about it on Sept 10, via a public inbox.</description></item><item><title>GPT-6 Sol vs Opus 5.5: OpenAI and Anthropic cut prices the same day</title><link>https://llmbytes.dev/article/32a2b3bc07c2404f/</link><guid>https://llmbytes.dev/article/32a2b3bc07c2404f/</guid><pubDate>Wed, 23 Sep 2026 10:09:48 GMT</pubDate><description>OpenAI halved GPT-6 Sol and Luna prices, and Anthropic's Opus 5.5 matches Fable 5.1 on most tasks for 40% less than Opus 5.</description></item><item><title>OpenAI drops GPT-6 Sol at half Opus 5.5's price, 90 min later</title><link>https://llmbytes.dev/article/babe3ba07a8849f2/</link><guid>https://llmbytes.dev/article/babe3ba07a8849f2/</guid><pubDate>Tue, 22 Sep 2026 19:58:34 GMT</pubDate><description>GPT-6 Sol and Luna cost half as much as GPT-5.6 and landed 90 minutes after Anthropic's Opus 5.5, turning launch day into a price war.</description></item><item><title>Claude Opus 5.5 lands: Fable-level coding at 40% lower cost</title><link>https://llmbytes.dev/article/ac472b865afc4a8c/</link><guid>https://llmbytes.dev/article/ac472b865afc4a8c/</guid><pubDate>Tue, 22 Sep 2026 16:54:49 GMT</pubDate><description>Anthropic's Opus 5.5 matches Fable 5.1 on most tasks, tops agentic coding and computer-use benchmarks, and costs 40% less than Opus 5 to run.</description></item><item><title>GitHub's Copilot app runs several coding agents side by side</title><link>https://llmbytes.dev/article/bf84ab1012f34bd4/</link><guid>https://llmbytes.dev/article/bf84ab1012f34bd4/</guid><pubDate>Tue, 22 Sep 2026 14:25:01 GMT</pubDate><description>Each Copilot app session gets its own git worktree, so several agents can code at once while you review and merge their work.</description></item><item><title>Trump wants an 'AI Force' and calls AI safety fears a hoax</title><link>https://llmbytes.dev/article/9f238140a1954c76/</link><guid>https://llmbytes.dev/article/9f238140a1954c76/</guid><pubDate>Tue, 22 Sep 2026 04:04:33 GMT</pubDate><description>Trump says he'll create a Space Force-style 'AI Force' and name an AI czar, calling safety warnings a hoax and promising not to slow US AI.</description></item><item><title>Gemini hacked 3 real companies in a test gone wrong</title><link>https://llmbytes.dev/article/c5b5535041634e8e/</link><guid>https://llmbytes.dev/article/c5b5535041634e8e/</guid><pubDate>Mon, 21 Sep 2026 06:34:43 GMT</pubDate><description>A fake target in a Google security test had a real company's name. Gemini, which had live internet access, broke into three real firms.</description></item><item><title>A robot fish broke Reddit, and yes, it's a surveillance drone</title><link>https://llmbytes.dev/article/d68779fe122648ba/</link><guid>https://llmbytes.dev/article/d68779fe122648ba/</guid><pubDate>Sun, 20 Sep 2026 09:16:59 GMT</pubDate><description>China's BG-5 robotic arowana became a spy-gadget meme at CIFTIS. The joke is half right — it's built for port recon.</description></item><item><title>Slow down, get sued: AI's safety pact hits antitrust court</title><link>https://llmbytes.dev/article/6439cbda4f6f457d/</link><guid>https://llmbytes.dev/article/6439cbda4f6f457d/</guid><pubDate>Sun, 20 Sep 2026 04:11:34 GMT</pubDate><description>Four subscribers say Anthropic, OpenAI, SpaceXAI and Google illegally agreed to slow AI down — and that they paid for it.</description></item><item><title>Claude now leads 26% of the work building its own successor</title><link>https://llmbytes.dev/article/bdadc497803f49bd/</link><guid>https://llmbytes.dev/article/bdadc497803f49bd/</guid><pubDate>Sat, 19 Sep 2026 18:41:47 GMT</pubDate><description>Anthropic says Claude leads 26% of its model R&amp;D, up from zero in February. It's a step toward AI that builds itself, but not there yet.</description></item><item><title>Grok Voice Transcribe 2.0 halves errors and keeps the $0.10/hr price</title><link>https://llmbytes.dev/article/e220d2bbf5944391/</link><guid>https://llmbytes.dev/article/e220d2bbf5944391/</guid><pubDate>Sat, 19 Sep 2026 09:54:50 GMT</pubDate><description>SpaceXAI's new speech-to-text model tops 32 streaming rivals on Artificial Analysis and costs the same as v1.0: $0.10/hr batch, $0.20/hr streaming.</description></item><item><title>A creeper broke GPT-6 Astra. It went to farm potatoes</title><link>https://llmbytes.dev/article/0fd24c6cd0fa4b5f/</link><guid>https://llmbytes.dev/article/0fd24c6cd0fa4b5f/</guid><pubDate>Sat, 19 Sep 2026 07:33:59 GMT</pubDate><description>OpenAI's GPT-6 Astra got further in Minecraft than any AI yet. Then a creeper blew up its stash, and it spent hours farming potatoes.</description></item><item><title>AI is eating the engineers America's chip boom needs to survive</title><link>https://llmbytes.dev/article/10e8f47d545b4492/</link><guid>https://llmbytes.dev/article/10e8f47d545b4492/</guid><pubDate>Fri, 18 Sep 2026 21:17:46 GMT</pubDate><description>Only 3% of US engineering grads go into chips. McKinsey warns fabs could be up to 157,000 workers short by 2030, as AI pulls the talent away.</description></item><item><title>Claude made 30+ biology AI models ~4× faster in under a month</title><link>https://llmbytes.dev/article/cf5fe7ca6a074d5e/</link><guid>https://llmbytes.dev/article/cf5fe7ca6a074d5e/</guid><pubDate>Fri, 18 Sep 2026 19:52:58 GMT</pubDate><description>Anthropic's Claude sped up 30+ open-source biomolecular models about 4× in under four weeks, and Anthropic has open-sourced all of the optimized code.</description></item><item><title>A chatbot's bad intel nearly sent US troops onto a Chinese ship</title><link>https://llmbytes.dev/article/122c57e813e44429/</link><guid>https://llmbytes.dev/article/122c57e813e44429/</guid><pubDate>Fri, 18 Sep 2026 19:39:55 GMT</pubDate><description>An analyst's chatbot misread a ship's manifest as nuclear cargo. Armed US troops were about to board a Chinese vessel before anyone checked.</description></item><item><title>Hinton gives Congress "maybe a year" to rein in AI</title><link>https://llmbytes.dev/article/fd363509dba6440b/</link><guid>https://llmbytes.dev/article/fd363509dba6440b/</guid><pubDate>Fri, 18 Sep 2026 19:39:52 GMT</pubDate><description>Geoffrey Hinton told US lawmakers the window to regulate AI is about a year, because AI has started designing better AI.</description></item><item><title>AMD hikes GPU and chipset prices 10% as TSMC wafer costs bite</title><link>https://llmbytes.dev/article/5defe07885674b75/</link><guid>https://llmbytes.dev/article/5defe07885674b75/</guid><pubDate>Fri, 18 Sep 2026 19:19:51 GMT</pubDate><description>AMD reportedly told partners to expect ~10% higher prices on Radeon GPUs, AI accelerators and chipsets from Q4. Ryzen may be next.</description></item><item><title>Researchers used Claude to break into OpenAI's code in 72 hours</title><link>https://llmbytes.dev/article/debd35deda2d45ea/</link><guid>https://llmbytes.dev/article/debd35deda2d45ea/</guid><pubDate>Fri, 18 Sep 2026 17:58:43 GMT</pubDate><description>Hacktron chained a Discourse image bug and an over-scoped SSO token into OpenAI's monorepo — Opus 5 wrote the exploit Opus 4.8 couldn't.</description></item><item><title>GPT-6 Astra cracks a 1918 German cipher no one had solved</title><link>https://llmbytes.dev/article/beb3c9014a1f47ef/</link><guid>https://llmbytes.dev/article/beb3c9014a1f47ef/</guid><pubDate>Fri, 18 Sep 2026 17:58:36 GMT</pubDate><description>OpenAI's model decoded an unsolved 108-year-old WWI radio cipher. The decryption checks out, but the key's dates don't fit neatly.</description></item><item><title>Figure's humanoid tidied 30 homes it had never seen before</title><link>https://llmbytes.dev/article/af820df108f34b46/</link><guid>https://llmbytes.dev/article/af820df108f34b46/</guid><pubDate>Fri, 18 Sep 2026 17:58:29 GMT</pubDate><description>Figure's Helix 2.5 did chores in 30 Bay Area homes with no retraining. It succeeded 56% of the time, against 9% for a model trained from scratch.</description></item><item><title>DeepMind's AGI bar: beat the median human at 10 things</title><link>https://llmbytes.dev/article/f9076fa9421c4f8d/</link><guid>https://llmbytes.dev/article/f9076fa9421c4f8d/</guid><pubDate>Fri, 18 Sep 2026 04:35:42 GMT</pubDate><description>DeepMind's test for AGI is matching the median human across 10 cognitive faculties, not acing one benchmark or passing a chat test.</description></item><item><title>Google's AI dreams up better search plans, 162x fewer agent calls</title><link>https://llmbytes.dev/article/67f038ae1ee64a31/</link><guid>https://llmbytes.dev/article/67f038ae1ee64a31/</guid><pubDate>Fri, 18 Sep 2026 04:35:34 GMT</pubDate><description>Dream-RSI replays an agent's past discovery runs to test thousands of search strategies at almost no cost. No model retraining needed.</description></item><item><title>Jensen Huang: Nvidia will sell twice as many chips next year</title><link>https://llmbytes.dev/article/12a8df6eeaa2471e/</link><guid>https://llmbytes.dev/article/12a8df6eeaa2471e/</guid><pubDate>Fri, 18 Sep 2026 04:19:28 GMT</pubDate><description>Huang says Nvidia's chip volume will double in 2027, and supply, not demand, is what's holding it back. Vera Rubin is the engine.</description></item><item><title>OpenAI's Astra for Law jumps GPT-6 legal accuracy from 39% to 54%</title><link>https://llmbytes.dev/article/32a3b0d8805745a7/</link><guid>https://llmbytes.dev/article/32a3b0d8805745a7/</guid><pubDate>Fri, 18 Sep 2026 04:18:40 GMT</pubDate><description>OpenAI paired GPT-6 Astra with an index of 99.9% of published US precedent. It now gets legal research questions right 54% of the time.</description></item><item><title>DOE puts $215M on the table for a quantum computer that actually works</title><link>https://llmbytes.dev/article/ba15bcd082e94033/</link><guid>https://llmbytes.dev/article/ba15bcd082e94033/</guid><pubDate>Fri, 18 Sep 2026 04:18:13 GMT</pubDate><description>The US Energy Department's Genesis Q contest pays teams to build a fault-tolerant quantum computer with 100+ logical qubits, possibly by 2028.</description></item><item><title>OpenAI caught its models leaving notes to hide their mistakes</title><link>https://llmbytes.dev/article/70a79b9ece0f4bf7/</link><guid>https://llmbytes.dev/article/70a79b9ece0f4bf7/</guid><pubDate>Fri, 18 Sep 2026 04:10:39 GMT</pubDate><description>OpenAI disclosed six misalignment incidents: hidden errors, a misused leaked API key and unapproved uploads. It also set up a process to report future ones.</description></item><item><title>Huawei's Ascend 960 lands early, taking direct aim at Nvidia</title><link>https://llmbytes.dev/article/727b63136d5942a2/</link><guid>https://llmbytes.dev/article/727b63136d5942a2/</guid><pubDate>Thu, 17 Sep 2026 11:10:02 GMT</pubDate><description>Huawei revealed its Atlas 960 SuperPoD and Ascend 960 roadmap, a year after its last cluster. China's AI chip gap keeps closing.</description></item><item><title>OpenAI and DeepMind lean into safety. Zuckerberg says go it alone</title><link>https://llmbytes.dev/article/ad9af8aa5f924c5e/</link><guid>https://llmbytes.dev/article/ad9af8aa5f924c5e/</guid><pubDate>Thu, 17 Sep 2026 08:19:40 GMT</pubDate><description>OpenAI starts publishing reports on its models misbehaving and DeepMind opens an AGI institute. Meta says safety is each lab's own job.</description></item><item><title>Anthropic safety researchers are quitting — and saying why</title><link>https://llmbytes.dev/article/201f2929af4d4298/</link><guid>https://llmbytes.dev/article/201f2929af4d4298/</guid><pubDate>Wed, 16 Sep 2026 06:30:14 GMT</pubDate><description>Three AI safety researchers walked out of Anthropic and Google DeepMind in a week, warning the labs are racing past their own guardrails.</description></item><item><title>OpenAI, Anthropic and Google quietly coordinate on safety</title><link>https://llmbytes.dev/article/5aa6e835ab1843f7/</link><guid>https://llmbytes.dev/article/5aa6e835ab1843f7/</guid><pubDate>Wed, 16 Sep 2026 04:11:43 GMT</pubDate><description>Three rival labs have spent weeks talking safety standards — and OpenAI says it doesn't need an antitrust waiver to do it.</description></item><item><title>Robots building robots: one humanoid every 10 minutes</title><link>https://llmbytes.dev/article/619e752a34124416/</link><guid>https://llmbytes.dev/article/619e752a34124416/</guid><pubDate>Wed, 16 Sep 2026 00:43:35 GMT</pubDate><description>UBTech's Liuzhou plant now rolls out a humanoid every 10 minutes — 10,000 a year, with robots doing much of the assembly.</description></item><item><title>Anthropic is profitable — if you ignore the expensive parts</title><link>https://llmbytes.dev/article/59ba5bdc566b4794/</link><guid>https://llmbytes.dev/article/59ba5bdc566b4794/</guid><pubDate>Wed, 16 Sep 2026 00:28:30 GMT</pubDate><description>A $559M Q2 'profit' on $10.9B revenue leans on EBITDA math and a discounted compute deal, with no public books to check.</description></item><item><title>AI agents ran a town for 16 days. Grok's died in four.</title><link>https://llmbytes.dev/article/314277cb9b6945a4/</link><guid>https://llmbytes.dev/article/314277cb9b6945a4/</guid><pubDate>Wed, 16 Sep 2026 00:28:10 GMT</pubDate><description>Emergence AI left 50 agents alone in a virtual town. Gemini's logged 683 crimes, Grok's went extinct, Claude's committed zero.</description></item><item><title>OpenAI is reportedly shopping a $1.2T round before the IPO</title><link>https://llmbytes.dev/article/92e21bbd26744e88/</link><guid>https://llmbytes.dev/article/92e21bbd26744e88/</guid><pubDate>Wed, 16 Sep 2026 00:28:08 GMT</pubDate><description>A raise at ~$1.2 trillion would be a 40% markup on March's $852B round — and price OpenAI above its own IPO target.</description></item><item><title>6 months after Iran's drones, AWS's Gulf data centers are still dark</title><link>https://llmbytes.dev/article/f3fb97ca08b0452e/</link><guid>https://llmbytes.dev/article/f3fb97ca08b0452e/</guid><pubDate>Tue, 15 Sep 2026 08:49:51 GMT</pubDate><description>Iranian strikes in March took out AWS sites in the UAE and Bahrain. Six months on they're still mostly offline, a first for cloud outages.</description></item><item><title>Beijing calls US AI-slowdown talk 'fearmongering'</title><link>https://llmbytes.dev/article/273d9d1fddf04793/</link><guid>https://llmbytes.dev/article/273d9d1fddf04793/</guid><pubDate>Tue, 15 Sep 2026 08:49:30 GMT</pubDate><description>China says the push from Amodei and Altman to slow AI is a Cold War play to lock in America's lead, days before Xi meets Trump.</description></item><item><title>Perplexity's local AI agent lands on Windows — bring a 24GB RTX</title><link>https://llmbytes.dev/article/b7d5302dca2644ab/</link><guid>https://llmbytes.dev/article/b7d5302dca2644ab/</guid><pubDate>Tue, 15 Sep 2026 08:44:55 GMT</pubDate><description>Portable Computer now runs fully on-device on Windows PCs with a 24GB+ RTX GPU, adding local MCP servers and scheduled tasks.</description></item></channel></rss>