Andrej Karpathy is a founding member of OpenAI and former head of AI at Tesla who now runs Eureka Labs. Opus 5.5 is Anthropic's Claude model, and Claude Code is its coding agent. Karpathy released the comment analysis behind HackerTube as a download so others can build their own versions.

Andrej Karpathy said on Nov 1 that, annoyed with YouTube's recommendations, he had Claude filter 43 million Hacker News comments from 20 years down to 6,454 videos that people keep recommending. The result is HackerTube, which he says took about two hours with Opus 5.5 and gives a significantly better viewing experience.

According to his post, the site needs no login and keeps its state in the local browser session. Thumbnails use a frame from the middle of each video instead of creator uploads tuned through A/B tests. He also released the underlying comment analysis as one zip file and suggested people point their own Claude at it to make their own version. The site also has a section for long videos.

Karpathy said he used Claude Code in the desktop app for Mac, with Opus 5.5 on Medium effort as his default, and only talked to Claude while it did everything. His only manual steps were in the browser: creating a Google Cloud API key to look up YouTube video data, starting a VPS for hosting and buying the domain. Once the server was up he gave Claude its IP address and told it to update, secure, harden, deploy and optimize the code, and he says it did the whole thing.

He wrote that many projects that would normally take weeks or months can now be done for breakfast. The post passed 1.3 million views and is one of the most viewed AI posts of the day.

Watch for people building their own versions from the released data, for example for other forums or other kinds of links, and for how well HackerTube holds up under heavy traffic now that the post has spread. The broader question is how much of a project like this an agent can handle without a person, since Karpathy still did the account, server and domain steps himself. Expect more builders to share similar end to end runs with Opus 5.5 and rival models.

SemiAnalysis is a research firm best known for its work on chips and AI infrastructure. GDPx is its new benchmark for agentic AI, built with outside data partners across 11 domains. The leaderboard image in the launch post averages each model across all domains, weighted equally.

SemiAnalysis (@SemiAnalysis_) announced GDPx on the evening of Nov 1 ET, describing it as a benchmark for measuring agentic capabilities across the most valuable domains. In its words, it is a true frontier level benchmark, since even the best model scores just 12%.

The leaderboard in the launch post ranks 13 of 20 models. Claude Opus 5.5 from Anthropic leads at 12.0%, followed by Google's Gemini 4 Carbon at 10.9%, OpenAI's GPT 6 Astra at 9.6% and Meta's Muse Wonder Full at 9.0%. Further down are Claude Fable 5.1, GPT 6.1 Sol, Claude Sonnet 5.5 and Gemini 4 Argon, with DeepSeek 4.1 Flash the best open model at 5.4%. Grok 4.7 scores 3.9% and GLM 5.3 2.7%. The chart also lists cost per task, from $0.92 for GPT 6 Luna to $67.16 for Claude Fable 5.1, with Gemini 4 Carbon at $16.29 against $51.41 for Opus 5.5.

In the thread, SemiAnalysis says each domain was built with specialized data partners, 14 in total, and that it did the final quality checks itself. Fireworks AI provided inference for open models and Modal provided sandboxes. It says GDPx focuses on domains with high return, such as chip design, cyber and bio, rather than weighting by share of GDP.

SemiAnalysis also says the scores of Opus 5.5 and GPT 6 Astra are held down by their safety classifiers, so it also reports results that exclude cyber, biology and alignment. It calls Muse Wonder Full the best available cyber model by a meaningful margin. Google CEO Sundar Pichai quoted the launch, pointing to the cost line for Carbon. These results come from SemiAnalysis and have not been independently reproduced.

Watch whether labs start citing GDPx scores in their own launches, and whether SemiAnalysis adds the other models in its full list. Its point about safety classifiers lowering scores could spark debate over how benchmarks should treat refusals. Expect the cost per task column to get as much attention as the scores, since Google is already pointing to it. Independent reruns of the tasks will show how stable the rankings are.

Chubby (@kimmonismus) is a widely followed AI news account that often shares leaks and early sightings. Mythos is the name of Anthropic's Claude model line referenced on the company's website. This claim is unverified: it rests on a screenshot sent to Chubby by an anonymous person, and Anthropic has not announced a Mythos 5.5 model.

Chubby, who posts on X as @kimmonismus, said on Nov 1 that Mythos 5.5 is already listed in the HTML source code of Anthropic's official website. He credited an anonymous person who sent him the tip by email. This is an unverified leak, and Anthropic has not announced or commented on a model with that name.

The screenshot attached to the post shows a browser viewing the source of the page anthropic.com/claude/mythos. In that source, a hidden page header and the alt text of a wordmark image both read Claude Claude Mythos 5.5. The screenshot does not show when it was taken, and its origin cannot be independently confirmed from the post.

Code on a company website can mention a product before launch, for example when a page is prepared ahead of an announcement, but it can also contain placeholder text. Seeing a name in page source does not tell readers when a model will ship, what it can do, or whether the name will stay the same.

The post drew more than 250,000 views within hours, making it one of the most discussed AI posts of the day. It landed on the same day as reports that OpenAI is testing a GPT 6.2 Astra model, which added to talk of a new round of frontier releases. Until Anthropic confirms details, Mythos 5.5 should be treated as Chubby's report of an anonymous tip.

Watch for an official announcement from Anthropic, a change to the Mythos page, or other people finding the same text in the site source. If the model is real, the key questions are when it launches, how it compares with Mythos 5 and with OpenAI's next models, and what it costs. If Anthropic removes or changes the text without comment, treat the leak with more caution until there is a clearer signal.

Token Gremlin is an X account that tracks AI model usage and pricing. GPT 6 Astra is the OpenAI model currently offered in Codex, OpenAI's coding agent. The claim of a GPT 6.2 Astra test is unverified and OpenAI has not announced the model. Chubby (@kimmonismus) amplified the report and later wrote that GPT 6.2 is very close.

Token Gremlin, an X account that follows AI model usage, claimed on Nov 1 that OpenAI is quietly A/B testing a new model variant called GPT 6.2 Astra. This is unverified. OpenAI has not announced a GPT 6.2 model or commented on the post.

According to the post, the Codex model picker still shows GPT 6 Astra, but the usage breakdown for some accounts lists GPT 6.2 Astra as a separate entry. Token Gremlin says multiple users are seeing the same thing and calls it a gray rollout, meaning a gradual test with a share of users. The attached chart shows one account's usage split as 80 percent GPT 6 Astra, 19.8 percent GPT 6.2 Astra and under 0.1 percent GPT 6 Luna. The post does not say whose account the chart comes from.

Chubby (@kimmonismus) quoted the report and later posted "Confirmed. GPT-6.2 very close," without naming a source. That post drew close to 80,000 views. Later in the day Chubby said he was seeing more posts from users who report GPT 6.2 Astra already rolled out in the Codex command line tool, and said he expects a release on Monday or Tuesday. Those are also his own expectations, not OpenAI statements.

The report came on the same day as an unverified leak about Anthropic's Claude Mythos 5.5, adding to talk that both labs are preparing new frontier models.

Watch for an OpenAI announcement or release notes for Codex naming GPT 6.2 Astra, and for more users reporting the model in their usage data or in the Codex command line tool. If it ships, the open questions are what changes against GPT 6 Astra, whether it reaches ChatGPT as well as Codex, and how it compares with Anthropic's next model. Chubby's expected Monday or Tuesday release is his prediction only.

Elon Musk owns xAI, which makes Grok, and X, which runs the X Money payments service. His post quotes Justine Moore (@venturetwins), an investor at Andreessen Horowitz, who shared a gift her Grok bot chose and paid for. Elon's post is the announcement; Justine's is the first hand demo.

Elon Musk wrote on Nov 1 that Grok's bot will research gifts your friends and family would love and buy them. He did not give a launch date, price or list of supported stores.

His post quotes Justine Moore (@venturetwins), who wrote on Oct 31 that a week earlier she had told her Grok bot to surprise her with a gift. She said it could spend up to $100 from her X Money account and asked it to use her posts and their conversations to pick something good. She then shared a photo of the gift that arrived, which she said even included a Halloween surprise. Her post had about 280,000 views and Elon's passed 660,000.

The example shows an assistant doing more than giving suggestions: it chose an item, paid with money from a linked account under a spending cap the user set, and had it delivered. That is one user's experience, and Elon's post describes a capability rather than a detailed product release.

The posts do not say how the bot confirms a purchase, what happens if it picks the wrong thing, how returns work, or which retailers it can buy from. They also do not say whether the feature is open to all users or limited to a test group.

Watch for xAI or X to publish details on the gift feature, including spending limits, how users approve purchases, refunds and which countries and stores are supported. Holiday shopping season will be the first big test of whether people trust an AI to spend their money. Also watch whether other assistants with payment features answer with similar shopping tools, and how retailers respond to orders placed by bots.

Grok Bot is an AI agent product sold on monthly plans with usage limits that reset on a schedule. On Nov 1, X user @DogecoinNorway posted that their usage hit 100% four days before the reset, then asked Elon Musk whether a reset was possible because many people said their usage ran out fast that week. Musk replied with the post above.

On Nov 1, X user @DogecoinNorway posted that their Grok Bot usage was at 100% with four days left until the reset. A few hours later the same account asked Elon Musk directly whether a reset was possible, saying many people reported their usage ran out really fast that week. Both posts passed one million views.

Musk answered in a short reply: the cost of serving Grok Bot was unsustainably high relative to revenue. The reply drew about 4.8 million views, making it one of the most viewed AI posts of the day. He did not say in that post whether limits would change again or whether a reset would happen.

An X news summary of the discussion says subscribers on plans from $30 to $200 a month reported their meters jumping quickly, and that Musk asked whether the higher tiers felt too expensive. That summary is generated from posts on X and has not been officially confirmed. The exchange turned a routine complaint about limits into a rare statement from Musk about the economics of running an AI agent at scale.

Watch for the company to explain the new limits, offer a reset, or add pricing tiers, since Musk's reply suggests the economics drove the change. Users will be comparing Grok Bot's limits with rival agent products from OpenAI and Anthropic. The broader question is whether always running AI agents can be priced at a flat monthly fee at all, or whether usage based pricing becomes the norm.

Anthropic disclosed on Oct 27 that it had acquired Mithril Medicine in July. Founder Cody Tranbarger, who posted on X for years as @Prof_Oak_, confirmed the deal himself and said he joined Anthropic to continue the programs and lead its broader internal drug development work. Martin Varsavsky, the entrepreneur behind Viatel and Jazztel, wrote about the deal in Spanish on Nov 1.

Martin Varsavsky (@martinvars) wrote in Spanish on Nov 1 that Anthropic, the company behind Claude, has bought a biotech that had a single employee. The deal itself is confirmed: Anthropic disclosed on Oct 27 that it acquired Mithril Medicine in July, and founder Cody Tranbarger announced it on X and LinkedIn.

According to Varsavsky, Tranbarger founded Mithril 18 months ago after starting as an analyst at Goldman Sachs and studying biotechnology and business at Harvard. Mithril is working on an antibody for LAMA2 congenital muscular dystrophy, a rare disease that appears in babies and children. Varsavsky writes that it had no lab of its own, only eight part time consultants, 11 outside labs and Claude coordinating contracts, logistics and experiments, and that it went from an idea to promising mouse results in under a year and a half. He says Anthropic hopes it could reach human trials by late 2028.

Tranbarger's own account matches the outline. He wrote that with a single full time employee, consultants, a network of contract research labs and a whole lot of Claude, Mithril took a first in class idea through proof of concept in animals in less than 18 months. He calls the model Pipeline in a Person and says Mithril's programs became the first in Anthropic's internal pipeline. R&D World reports a second program targets trials as early as 2029.

Varsavsky argues the bigger point is cost: rare diseases with few patients rarely justified big teams, and he suggests the next big biotech may have one employee. That is his view.

Watch for Anthropic to share more on its internal drug pipeline, including whether it starts more programs run by single operators and whether it offers the approach to outside biotech companies. The key scientific test is whether the LAMA2 antibody keeps working in larger studies and reaches human trials on the late 2028 timeline Anthropic has given. Regulators and investors will also look at how much of the work Claude did versus the consultants and contract labs.

The US Bureau of Economic Analysis released its advance estimate of third quarter 2026 GDP on Oct 29. It reported real GDP growth at a 3.4% annual rate, with the rise in equipment investment led by information processing equipment, mainly computers and peripherals. The BEA does not publish a separate category for AI spending, so figures like these are a proxy for the AI buildout.

The Kobeissi Letter (@KobeissiLetter) wrote on Nov 1 that AI spending is rapidly accelerating in the US. It says private fixed investment in information processing equipment and software grew 7.3% in Q3 2026 to a record $1.77 trillion, the largest quarterly increase since Q4 1983, compared with 6.2% at the height of the 2000 dot com bubble. It adds that the measure is up 61%, or $671 billion, since Q4 2022, and has reached a record 5.4% of GDP versus a previous high of 4.5% in Q4 2000. The post calls this annualized growth, while its own chart labels the 7.3% as quarter over quarter growth. These calculations are the account's own and were not checked line by line here.

The underlying data comes from the BEA advance estimate released Oct 29. BEA reported real GDP growth of 3.4% at an annual rate and said the gain in equipment was led by information processing equipment, mainly computers and peripherals. Coverage of the release reports that real investment in information processing equipment rose at a 37.4% annual rate and software at 9.6%. PNC economists estimate that information processing equipment, software and research together added 1.30 percentage points to growth, nearly double the second quarter.

There are caveats. BEA based the equipment figure mostly on import data, and much of the hardware is made abroad, so imports offset part of the boost to US output. The advance estimate is also the first of three and can be revised.

BEA publishes its second estimate of Q3 GDP on Nov 25, which will show whether the record holds after revisions. Watch the fourth quarter data for signs that AI equipment spending is still accelerating or starting to level off, and whether more of the hardware is produced in the US. Investors will also compare these totals with the capital spending plans that large tech companies give in their earnings calls.

Power has become one of the main limits on building AI data centers in the US. The zerohedge article draws on two Wall Street reports: a Goldman Sachs US Power Pipeline report built on Energy Information Administration data, and an estimate by Morgan Stanley analyst Stephen Byrd. The figures below come from zerohedge's account of those reports, which are not public.

Zerohedge (@zerohedge) posted on Nov 1 a link to its article titled The 70 Gigawatt Hole: Wall Street Maps How Far America's Electric Grid Is Falling Behind AI Demand. The post itself only gives the headline, so the details come from the article, which in turn summarizes bank research.

According to zerohedge, the grid had a strong month on paper. It cites Goldman Sachs' utilities team, working from Energy Information Administration data, as saying US generating capacity rose by a net 5.6 gigawatts in September. Zerohedge says that was the strongest September in at least 10 years and almost four times the 1.5 gigawatt historical average.

In the same week, the article says, Morgan Stanley's Stephen Byrd calculated that US data center developers face a power shortfall of 70 gigawatts through 2028 before counting stopgap sources. Those include turbines, fuel cells and converted bitcoin mines. Even after counting them, zerohedge reports, the gap is still 43 gigawatts, or 38% of all the power AI developers need. The link preview quotes Morgan Stanley saying power availability continues to be the most important constraint, and that it is most notable in the US.

The bank reports themselves are not public, so these figures are attributed to zerohedge's summary and were not checked against the originals. The story fits a wider pattern in this deck of AI growth running into physical limits.

Watch for new utility filings, gas turbine orders and grid connection rules that show how much of the shortfall can be closed before 2028. Data center builders are likely to keep turning to onsite generation, fuel cells and power purchase deals, which could raise costs. Regulators and lawmakers may also face pressure to speed up transmission approvals as the gap becomes a limit on AI growth.

The AI boom needs huge new data centers, and many towns are pushing back over noise, water use, power prices and land. The figure comes from a Gallup survey, shared on X by unusual_whales. The Gallup report itself was not linked in the post.

The markets news account @unusual_whales posted on Nov 1 that 71% of Americans opposed the construction of a data center in their local area, citing Gallup. The post drew more than 250,000 views, more than 3,000 likes and over 150 replies by the time of this scan.

An X news summary of the poll says opposition to local data centers was higher than opposition to a local nuclear plant, which it puts at 53%, That summary is generated from posts on X, and the full Gallup release was not checked for this story.

The poll lands in the same week that, according to zerohedge, Morgan Stanley estimated a 70 gigawatt US power shortfall for data centers through 2028, a story already in this deck. According to Axios, OpenAI is planning a campaign right after the midterms to rebrand data centers as AI compute campuses, which suggests the industry sees public opinion as a growing problem. Taken together, the poll and the power estimates show two separate limits on AI buildout: whether there is enough electricity, and whether communities will accept the buildings at all.

Watch whether local fights start blocking or delaying named projects, and whether developers offer community benefit agreements, lower power bills or local hiring to win approval. The Nov 3 midterm results may shape how much state and federal support these projects get. AI companies, led by OpenAI's planned campaign, are likely to push harder on messaging about jobs and tax revenue.

The US midterm elections are on Tuesday, Nov 3. Andrew Curran quoted his earlier post on a New York Times report that lawmakers signaled the House would move away from its largely hands off approach to AI companies if Democrats retook it. Chris Lehane is OpenAI's chief global affairs officer.

Andrew Curran (@AndrewCurran_) posted on Nov 1 that OpenAI is preparing for a post midterms change in the weather, sharing an excerpt from Axios. According to that excerpt, right after Tuesday's midterms OpenAI will launch a policy and PR blitz to get in sync with the new Washington, including an effort to rebrand data centers as AI compute campuses.

Axios writes that after two years of a friendly White House, AI companies will face a harsher Washington after Tuesday, when it says Democrats will likely win the House. It adds that OpenAI also aims to help frame the AI debate for the field of 2028 presidential contenders. OpenAI chief global affairs officer Chris Lehane is quoted saying you can try to chase the moment or seek to inform it through ideas, and calling Tuesday an electoral portal from one era to the next.

Curran's post quotes his own earlier post on a New York Times report, which said lawmakers signaled the House would likely shift away from its hands off approach to AI companies if Democrats retook it.

An X news summary adds that the campaign would start in the lame duck session from Nov 9 and include ideas such as speed limits on powerful systems and community agreements for projects. Those details come from the summary and were not confirmed in the Axios excerpt.

Watch the Nov 3 results and whether Democrats take the House, which would change who leads AI oversight hearings. OpenAI's campaign is expected to start soon after, so look for new policy papers, ads and the AI compute campuses branding. Rival labs may respond with their own pitches, and data center opposition will likely be a central theme.

Polymarket is a prediction market whose X account posts news headlines. The underlying report is a Fortune article from Oct 30 in which Cisco President and Chief Product Officer Jeetu Patel described the project. The engineer is Padmanabhan Krishnamurthy, 27, who joined Cisco in 2025 when it bought his translation startup EzDubs.

Polymarket posted on Nov 1 that Cisco's Chief Product Officer gave a Gen Z engineer unlimited AI tokens to vibecode a rewrite of 98% of the company's legacy code within 2 years. The post is a secondhand summary, and the word vibecode is Polymarket's.

The source is a Fortune report from Oct 30. Fortune says Cisco President and Chief Product Officer Jeetu Patel told it at a small roundtable dinner that he had given unlimited tokens to Padmanabhan Krishnamurthy, known as Paddy, because he is going to refactor code, and told him he could go nuts. Krishnamurthy is 27 and joined Cisco in 2025 after it acquired EzDubs, the real time translation startup he cofounded in 2023. Patel said he had been buried in the organization, five levels below him.

According to Fortune, Krishnamurthy said he could rewrite Cisco's legacy code with a team of two, and has given himself four weeks to prove the approach, with the final goal of refactoring 98% of the legacy code within 24 months. Patel said even the most AI forward people on his team were not thinking that way.

India Today later reported that Patel wrote on X that the two to three person team is meant to start the effort and have Cisco's core engineers own the final product, that it could help move more code to memory safe languages, and that not every experiment will work but this one is worth pursuing.

Watch for whether Krishnamurthy's four week proof period produces results Cisco shares publicly, and whether the company sets out how it will test and review AI rewritten code in products that run networks around the world. The 98% goal over 24 months is a target, not a result. Other large companies with old codebases will be watching to see if small AI heavy teams can deliver this kind of rewrite safely.

Mythos is the name of Anthropic's top Claude model line. The repo, burner22/numbat on Hugging Face, claims to contain an internal training checkpoint, but the poster himself stresses that nobody has verified the weights and Anthropic has not commented. This is a separate claim from the leak of a Mythos 5.5 name in Anthropic's site code, also in today's deck.

Shams (@OxShamss) posted on Nov 1 that an alleged Mythos class model with 14 trillion parameters had appeared on Hugging Face, and in the same post said the problem is that nobody has verified the weights. He asked whether it was a real leak or an elaborate fake.

According to the post, the repository claims 14 trillion parameters, 1,224 mixture of experts components, 7.72 TB of weights split across 772 shards, a hybrid GDN plus MLA architecture, and model configs, architecture code and training metadata. The attached screenshot names the repository as burner22/numbat and says it claims to contain an internal Anthropic training checkpoint, with NVFP4 quantization. All of these figures are the repository's own claims.

No one quoted in the post has loaded the files or tied them to Anthropic, and Anthropic has not commented. A repository can describe itself in any way, and large file sizes or plausible configs do not prove authenticity, a point the poster makes himself. The account name burner22 also gives no clue to who uploaded it.

The post drew close to 280,000 views on a day already busy with Anthropic rumors, including a separate unverified report that the name Mythos 5.5 appears in the HTML source of Anthropic's website. The two claims are unrelated events and neither has been confirmed.

Watch for independent researchers to inspect the files, check whether the weights load and behave like a frontier model, and compare the architecture with anything Anthropic has published. A takedown request from Anthropic or removal by Hugging Face would be a strong signal either way. Until someone verifies the contents, treat the repo as an unconfirmed claim and be cautious about downloading or running unknown files.