Arena

Tesla is hiring a Machine Learning Engineer for Realtime Games, Digital Optimus, in Palo Alto. The job: build agents that play AAA games, browse the web and use pro apps at superhuman levels. Why games? Digital Optimus, the Tesla and xAI project Musk announced in March, reads screen video and acts with a keyboard and mouse. AAA games are a safe, endless, real-time proving ground for exactly that, the same approach DeepMind's SIMA uses. Tesla's posting says it will benchmark how skills transfer across games, general computer use, FSD and Optimus. tesla.com/careers/search/job…

$LMND posting strong growth again this quarter, and a 64% net loss ratio. I’d say the AI is working. Affirmed positive Adj EBITDA guidance for Q4, which is a huge milestone. I’ll keep digging but this looks very solid.

MICRON $MU PAUSES CONSTRUCTION OF UP TO $4.5 BILLION MEMORY CHIP PLANT IN TAIWAN AMID LABOR DISPUTE. The Taoyuan facility was expected to begin producing DRAM and HBM chips in 2028. Workers are demanding higher compensation, including a proposal to distribute 15% of Micron's operating profits to employees globally. Micron says existing Taiwan operations continue as normal. The island accounts for roughly 60% of the company's global DRAM production.

ACM Research, $ACMR, Q3-26. Sales crushed estimates and guidance jumped. An investment loss hit GAAP earnings. 🟢 Revenue: $367.7M vs. $310.6M est. 🟢 Adj. EPS: $0.55 vs. $0.51 est. 📈 FY26 guide: $1.24B–$1.28B (raised) ⚠️ GAAP EPS: $0.02; $39.1M unrealized loss

ERNIE 5.5 Flash Base is now open source — a 280B-parameter model with 14B active per token, and it already matches SOTA at its size! And this is just the start — the full ERNIE 5.5 Flash is on its way. 👀

Hy4 from Tencent scores 35 on the Artificial Analysis Intelligence Index v5 and is the #2 model from a China-based AI lab Hy4 is the latest model from @TencentHunyuan and is the successor to Hy3 released earlier this year. It scores 35 on the Artificial Analysis Intelligence Index v5, equal with Grok 4.7 (xhigh) and 1 point behind Claude Haiku 5.5 (max). Hy4 is the #2 model from a China-based AI lab, behind only Kimi K3.1 Pro, which scores 46. Hy4 is priced at $0.834/$2.501 per 1M input/output tokens. Key takeaways: ➤ Hy4 excels at agentic coding tasks. Hy4 scores 31% on FrontierCode 2.0, compared to 32% for Claude Fable 5.1 (max). It is the leading model from a China-based AI lab, with DeepSeek V4.1 Flash 1017 (max) scoring 28% and Kimi K3.1 Pro scoring 25%. On Terminal-Bench 4.0, it scores 37%, the fourth-highest score among models from China-based AI labs ➤ Hy4 outperforms its peers on knowledge and reasoning. Hy4 achieves accuracy of 50% on AA-Omniscience v1.1, 1 point ahead of Kimi K3.1 Pro, but has a high hallucination rate of 89% compared to Kimi K3.1 Pro’s 37%. On HLE-Diamond, it scores 33% compared to Kimi K3.1 Pro’s 37% ➤ Hy4 offers a competitive Capability vs. Cost trade-off compared to to peers. With a Cost per Intelligence Index Task of $0.78, Hy4 (35) is 10-times cheaper than Grok 4.7 (xhigh, 35) and 13-times cheaper than Qwen3.8 Max (0902) (34) but almost twice as expensive as Claude Haiku 5.5 (max, 36) Other model details: ➤ Context window: 1 million tokens ➤ Pricing: $0.834/$2.501 per 1M input/output tokens. Cache reads $0.042 per 1M tokens ➤ Multimodality: Text and image input, with text output

AI is using more and more of the world’s energy. But how much economic value are we getting from AI? We’re introducing Joules-to-GDP (J2G), a framework that measures how well AI turns energy into economic value and shows which tasks, models, and hardware give the best return. Across 44 occupations in finance, engineering, law, administrative work, and more in @OpenAI's GDPval, we arrive at three findings: 1) The tasks with the highest economic value are not the tasks with the highest economic value per joule. For example, CAD design earns ~10% more per task than building websites, but the latter earns ~27% more value per joule. 2) The most powerful model isn’t always the best use of energy. Gemma 4 31B delivers 56% of GLM 5.2’s net value with 23% of the energy, for 2.5x the value per joule. 3) More chips don’t guarantee proportionate returns. Moving the same model from 1 to 8 NVIDIA H100s uses 3.2x the energy but cuts completion time by only 22%, with about the same value. With @SnorkelAI, we’re also releasing GDPWorks, a benchmark for economically valuable tasks across 36 additional occupations. Joint work co-led by @amirzeinaalii and @Avanika15, with @Mersad_Abbasi, @Jon_Hartley_, John Hennessy, @HazyResearch, and @Azaliamirh. More details below 🧵

OpenAI’s Greg Brockman reveals an unannounced JPMorgan Chase deal, with the bank using Codex and OpenAI’s cyber models to secure itself “Well, so within the company, we were probably, I don’t know. Like, I think our enterprise revenue was about 30% versus consumer at 70%. And now we’re probably somewhere around, like, we’re majority enterprise. And it’s not that consumer has shrank. Right? It’s been growing, but it’s the enterprise segment has really taken off." "And so I don’t know exactly how to quantify in the broader market, but you can see the way that we’re winning deals. So for example, actually, one that’s hot off the press, has not been previously announced, is that we just papered a partnership with JPMorgan Chase, and that they’re going to be using Codex." "They’re going to be using our cyber models to secure themselves, deploying our models for enterprise transformation knowledge work across the company." "And that for me is like such an iconic institution in such an important domain. Right? A highly regulated enterprise that trust is, like, so key to what they do and that they chose us and that we’re here to be a great partner.”

After 1 year, Dhruv Rathee just revealed the numbers behind his AI startup, AI Fiesta: - 1,00,000+ paying users - 9 crore+ messages sent - 27 team members - 0 external investors - ₹47 crore in total revenue

Kimi-K3.1-Pro by Moonshot AI: interesting facts It's huge, and sparse • It has 2.8T total parameters, but only 104B are active per token. • Along with Kimi-K3, it's the largest real model release on the Hub. The next biggest are Qwen3.8-2.4T and DeepSeek-V4-Pro (1.6T). • It has 896 experts and sends each token to 16 of them. (DeepSeek-V4-Pro has 2 times less and Mistral-Large-4 has 7 times less) Clever attention for a 1M-token context • Most of its 93 layers use linear attention (KDA). Every 4th layer uses DeepSeek-style sparse attention, which only looks at the top 2048 tokens. • This is how it handles a 1M-token context (the same as the other frontier open models) • It uses a custom activation called SiTU-GLU and has a 401M-parameter vision encoder (MoonViT-V2), so it takes both text and images. Small on disk for its size • It was trained to run in 4-bit (MXFP4), so the whole repo is only ~1.56 TB. • Stored in regular BF16 it would be about 5.6 TB(!) Benchmarks (from the model card) • It goes back and forth with Claude Opus 5.5 and GPT-6 Astra. It's top of the card's table on BrowseComp (92.1), HLE with tools (70.0) and SWE Marathon (51.7). • It's weaker on pure reasoning without tools and on cyber security. • Moonshot ran most of their own coding evals using the Claude Code harness. License (a modified MIT) • It's free to use, modify and redistribute. • API businesses making over $20M a year need a deal with Moonshot. • Products with over 100M monthly users or over $20M monthly revenue must display "Kimi K3.1-Pro". On the Hub • It went public about 1.5 hours ago and already has 214 likes, about 9× Kimi-K3's first hour. It's #19 on the trending list. • It isn't gated, but loading it in transformers requires "trust_remote_code"

Finpilot Raises $10 Million Series A Led by Lobby Capital to Bring AI Agents to Institutional Investors ow.ly/lnKl5103ZVE #financialtechnology #Fintech #financial #finance #InsurTech #FintechNews #AIinFinance

Nooks, the AI Revenue Agent Platform for human + AI collaboration across the full revenue lifecycle (prospecting to renewals), secures $500M from General Catalyst’s Customer Value Fund—non-dilutive capital where GC takes a share of the returns its spend generates.

AI inference startup fal raised $1.3 billion in a Series F at a $16 billion valuation, Bloomberg and TechCrunch report. Bessemer led the round, with Nvidia, Sound Ventures, HOF Capital and T. Rowe Price joining. thecircuitry.to/article/fal-…

Your Pro API credits do not buy you more Codex. OpenAI wrote that in plain English. And starting in December, ChatGPT and Codex share one bar anyway. If you pay for ChatGPT Pro, you may have noticed a new line on the billing page: monthly API credits. On the $100 Pro tier you get $100. On the $200 tier you get $200. On the $500 tier you get $500. Business seats get $20 per Standard seat or $100 per Premium seat, pooled under a $500 per month workspace cap. Those credits land on an API organization you own. They cover paid API Platform work on that org, including the Agents API, Responses, Images, and related tools. Unused credits expire at the end of the billing cycle. They do not roll over. When the free balance is gone and you have not bought more, requests fail with HTTP 429 and the error code credit_balance_exhausted. You also need the subscription to be active for seven days before you can claim. Plus, Go, and Free are not eligible. None of that buys you extra Codex. OpenAI's own developer docs say Codex usage through API keys is not supported by this benefit. The same paragraph adds that these API credits also do not increase your ChatGPT or signed-in Codex usage limits. That is the whole trap in one sentence. People hear that Pro now includes hundreds of dollars of API credits and assume the coding agent sitting inside the same account just got a bigger tank. It did not. The credits are for building apps against the API org you linked. Signed-in Codex still burns the plan allowance. ChatGPT Work still burns the plan allowance. The shiny API grant sits beside both of them and does not refill either meter. Then comes December. OpenAI's Help Center article on ChatGPT and Codex usage limits says that starting in December, Plus, Pro, and Business plans will share one usage allowance across ChatGPT and Codex. The reason they give is that they are bringing the agentic capabilities of Work into the main ChatGPT experience. Usage will be measured in tokens. How much you spend depends on the model you pick and the task you ask. Instant chat will not draw from that shared pool. You will see combined remaining usage and reset times under Settings, then Usage. If you hit the limit, you can wait for a reset, upgrade your plan, or buy credits. OpenAI says it is preparing carefully for the transition and that it expects most users will not need to wait for a reset more often than they do now. That line is an expectation, not a published token budget. The help article still does not list how many tokens each plan gets. Developers noticed. On October 31, a community feature request titled "Please Keep ChatGPT and Codex Usage Limits Separate in December 2026" went up under the handle Smalikou. Through November 1 and 2, dozens of Plus and Pro users piled on with the same complaint: chat and coding already fight for attention in a workday, and under one pool they will fight for the same remaining bar. OpenAI has not answered that thread with exact numbers. Until those numbers appear, the only hard facts are the ones on the help page. One shared allowance. Token metering. Instant excluded. Three exits when you hit the wall. Put the two changes next to each other and the product story gets clearer. OpenAI added API credits that look like a gift to builders, then clarified that Codex is outside that gift. Next it will fold ChatGPT and Codex into one bar so a long agent run and a long chat session pull from the same wallet. The marketing line is convenience. The billing line is consolidation. If you use Instant for quick questions, you are fine. If you use Work for multi-step jobs and Codex for repo work on the same day, December is the month those two habits start taxing each other. Anthropic is running a different shape of the same problem. On Claude's pricing page this week, Max is a single $100 tier with 5x Pro usage every five hours and every week, plus $100 a month in API credits. The old Max 20x plan is now Ultra at $200: 10x Pro weekly usage, no five-hour limit for verified accounts, and $200 a month in API credits. Existing Max 20x subscribers move to Ultra automatically with no price change. Claude's FAQ is blunt about the pool. Activity across Claude on the web, desktop, and mobile, and Claude Code, draws from the same allowance. Claude Code is included in the paid plans. So Anthropic also makes chat and the coding agent share a meter. The packaging is just clearer about it. Ultra removes the short window that used to pause heavy users every five hours, then leaves the weekly ceiling in place. OpenAI is about to make the shared-meter story explicit for ChatGPT and Codex while selling API credits that, by design, do not refill that meter. Here is what you can verify today without guessing. Pro and Business API credits are real, tiered, and expire. They do not cover Codex through API keys. They do not raise ChatGPT or signed-in Codex limits. From December, Plus, Pro, and Business get one shared ChatGPT plus Codex allowance measured in tokens. Instant chat stays outside that allowance. OpenAI has not published the token sizes yet. A community thread started October 31 is already asking to keep the limits separate. Claude Max and Ultra also put Claude Code on the same pool as the chat products, with Ultra dropping the five-hour pause for verified accounts while keeping a 10x weekly cap. What you should not invent is a personal burn rate, a December day number that OpenAI has not named, or a claim that API credits suddenly buy Codex sessions. They do not. OpenAI wrote that down. Read the credits page before you plan a week of agent work around a number that only applies to the API org. If you are on Pro mainly for Codex, treat the API credit line as a separate product. Build with it if you want. Do not count it toward your coding headroom. If you are on Plus and lean on both ChatGPT Work and Codex, watch Settings then Usage once the shared bar lands. When the bar hits empty, the exits are wait, upgrade, or buy credits. That is the whole menu. The interesting failure mode here is a quiet merge of two products that used to feel like separate budgets, right after a credit grant that looked like it solved the problem and explicitly does not.

Trump Media just moved $344 million in Bitcoin to an exchange. Wallets linked to the company sent 4,259 BTC to Crypto.com on Wednesday, per on-chain data flagged by Lookonchain and Arkham. The amount nearly matches the 4,260.73 BTC the company disclosed as collateral for $1 billion in convertible notes. The same day, its shares fell 24.5% to a record low after the midterm results. A transfer to an exchange is not proof of a sale — and Crypto.com is one of its custodians. But with noteholders able to demand cash repayment on November 30, the timing is the story. Bitcoin trades near $80,400.

BREAKING 🔥: GPT-6.2 Astra and GPT-6.2 Sol have been spotted on the Google Cloud Console quotas page! They are indeed coming soon! 👀

Images, upgraded—introducing ChatGPT Images 2.6. Now with GIFs and reference images. Plus, two new models are coming to the API with 50% faster image generation.

Sam Altman needs ChatGPT to answer Raj Shamani's questions ! 😭

$LITE $AAOI $SIVE $AXTI There it is! $COHR just confirmed the AI optics supercycle is stronger than ever. > Revenue $2.37B (+50% YoY). > EPS $2.09 (+80% YoY). > Datacenter & Comms: $1.95B (+79% YoY). They just guided FY27 pro forma growth to >60% and a path to >45% gross margins. This is not a beat. This is acceleration. Copper is dead. Every new AI cluster needs more optics, and only a handful of companies can deliver at scale. > $LITE just got bought by $MRVL for $115B. > $AAOI is the purest transceiver play. > $SIVE is the high-beta laser/CPO call option. > $AXTI is the InP bottleneck, everyone needs them. Demand is not slowing, it’s accelerating into 2027! The market is still debating “is AI capex sustainable?” $COHR just answered with a >60% growth guide. Position accordingly. Take advantage of this generational wealth creation opportunity. We remain long and bullish. NFA.

Google just finished 33rd out of 42 in a search contest. Microsoft came 30th, Amazon 29th, and OpenAI 25th. First place went to open-source code anyone can download and run on their own server. The contest covered the kind of search everyone at an actual job deals with: find the right answer somewhere inside the company's own Slack, email, Jira tickets, wiki pages and Drive folders. Onyx built it as EnterpriseRAG-Bench: 500 questions over 511,962 documents from a made-up company, junk left in on purpose, GPT-5.4 grading, and a wrong answer scores zero even if it sounds great. Top of the board right now is PipesHub at 93.24 out of 100. OpenAI File Search got 61.03, Amazon Q 48.96, and Google's Vertex AI Search 41.87. My read on why the giants struggle here is pretty boring. They ship the public-web recipe: turn documents into numbers, grab the closest ones, hand them to the AI. Company data breaks that. Picture asking "what's our refund policy" when the real answer lives in a Slack thread that overrides the wiki page, and the wiki still ranks first because it says "refund" forty times. Getting into the 90s means knowing which source wins a fight, and the small teams at the top look like they spent a year on little else. In May 2023 a Google memo leaked titled "We Have No Moat, And Neither Does OpenAI." The worry then was open-source models. The clearest open-source win since showed up in the part that finds the right document: 32 points over OpenAI, 51 over Google, from code anyone can download tonight. github.com/pipeshub-ai/pipes…

Neoclouds are CRASHING 10% after Texas turned blue and its new governor threatened to pause data center construction. $NBIS has ZERO MW in Texas. Not online. Not in the pipeline. NOTHING. Its U.S. pipeline is elsewhere: • Independence, Missouri: 1.2 GW • Pennsylvania: up to 1.2 GW • Birmingham, Alabama: 300 MW • Vineland, New Jersey: 300 MW And it has something no other neocloud has: a real European footprint. Finland, France, the UK, Iceland, Israel, Spain. If the U.S. gets hostile to data centers, $NBIS is the hedge. In fact, a pause in Texas, home to 27% of planned U.S. data center capacity, is BULLISH for $NBIS. Less new supply = every MW already online is worth MORE. The market is selling the wrong stock.