Welcome to Import AI, a newsletter about AI research. Import AI runs on arXiv, cappuccinos, and feedback from readers. If you’d like to support this, please subscribe. Subscribe now LLMs are very bad at low-latency tasks:…Realtime-Bench say…
Firehose
Filtered to People · clear filters
Browse: People · Companies · Papers · Podcasts · Hacker News · Deep dives
Browse by tag
I was reminded today of AI Adoption Rate Trending Down for Large Companies , a chart produced by Apollo Global Management in September 2025, which appeared to show a slowdown in U.S. corporate AI adoption based on numbers collected by the U…
Google's earnings miss is similar to Facebook's in terms of aggressiveness, indicating the company is being more proactive than previously thought. This mirrors Facebook's own more aggressive approach to growth and expansion. Google's earnings report shows the company is taking a more measured approach to growth, which is a departure from its typical more conservative approach. AI summary
I've been getting the occasional confusing Python dependency error since switching over to Python 3.15 and the rule I'm having to remember is: If in doubt, delete the uv.lock file Most of my projects are libraries which don't include a uv.l…
More broadly, the use of LLMs erodes trust in documentation, in software, and in the organizations which produce them. Perhaps engineers will begin to treat any new software project with suspicion—the signals we used to rely on as markers o…
How We Took Over ChatGPT Accounts Parameter Labs describe an exploit they found against ChatGPT's implementation of the artifacts pattern (OpenAI used to call this "Canvas" but it looks like the latest revision is called HTML previews ), wh…
Release: datasette-llm 0.1a9 I have some big upgrades to Datasette Agent in the pipeline, but in the meantime here's a tiny user experience improvement to the underlying datasette-llm plugin. It now picks up the default model used by LLM , …
The author created a YouTube clone, HackerTube, using Opus 5.5, a workflow automation tool, to scrape and filter 6,454 YouTube video recommendations from Hacker News comments, and then deployed the site on a virtual private server (VPS). The site uses a custom YouTube UI, provides good "ForYou" recommendations, and allows for keyboard shortcuts and data export. The entire process took around 2 hours, with Opus handling most of the work. AI summary
Arca, an AI-native wealth management firm, uses agents to perform long-horizon financial work, such as analyzing client data and generating financial plans that can take up to ten days to complete. The firm has developed an internal memory system to track the progress of these tasks and ensure that human advisors are involved in the decision-making process. Arca's agents are designed to prompt humans for input, rather than the other way around, and the firm has implemented various safety measures to prevent agents from making decisions that could harm clients. AI summary
The Pope's Encyclical on AI, "Magnifica Humanitas," emphasizes five key points for guiding the development of artificial intelligence: the future of AI is not predetermined, machines are tools not persons, certain decisions must never be delegated, intelligent technologies must serve the common good, and institutions are needed to put ethical principles into practice. AI summary
Google's AI Mode has reportedly led to a -15% drop in search revenue, equivalent to a $500B loss, the worst since its 2004 IPO, with the AI Bear Market potentially looming. AI summary
dbreunig/nesformer Inspired by qwenputer by Divyansh Agrawal, Drew Breunig baked a working emulator for a Nintendo Entertainment System 6502 CPU into the weights of Qwen3.5-4B. You can now play NES ROMs such as Super Mario Bros (very slowly…
Nikole Hannah-Jones wrote a brave self-critical piece in The New York Times, centrally about how she intentionally sent her daughter to a well-funded but failing school, PS 307, in the hopes of using voice to improve it, and how that utterl…
Last week coined the term “Black October“ after an onslaught of critical vulnerability disclosures, including PHP2Shell and GitLab, signified the moment when the cyber risks of AI models started to materialize.
The best Stratechery content from the week of October 26, 2026 including Anthropic numbers, the end of the Internet as we know it, and autocracies then and now.
The rapid advancement of AI technology has significantly increased the risk of software security breaches, making it essential to rethink software security measures. AI-powered attackers can now automate and speed up attacks, exploiting critical vulnerabilities before patches are released, and accessing systems externally or remotely. To combat this, automated patching and ongoing monitoring of security breaches will become non-negotiable, and defenders will need to use AI agents to stay ahead of attackers. AI summary
Anthropic's S-1 filing provides a detailed look at the company's financials, including a valuation of $2.5 billion, and offers insights into the competitive landscape of large language models, with implications for Meta, Muse, and other AI-focused companies. The filing also sheds light on Anthropic's approach to scaling its models and its plans for future growth. AI summary
The AIE CODE conference, the top Coding Agent conference, features three themes: Scaling without Slop, System One models, and Increasing Ambition, with a focus on showcasing the latest advancements in agentic coding and cybersecurity. AI summary
API credits benefit for eligible ChatGPT subscriptions Earlier this month Anthropic announced a generous new benefit for their Max and Team subscriptions: $100, $200, and $500 per month subscribers would get the equivalent quantity of (expi…
How I Built the Most AI-Pilled Company in the World Don't be put off by the clickbait headline, this piece by Jacob Sansbury describing how they are building Spawn is a fascinating exploration of the bleeding edge of what's possible with ag…
Elorian's Glimpse model has demonstrated 50.48% accuracy on the DecoyCount benchmark, surpassing frontier models like GPT-6 Astra (43.17%) and Gemini-3.8-flash (32.71%). This achievement highlights Glimpse's ability to visually reason and ground its counts in visible evidence, rather than relying on shortcuts or substitution heuristics. The DecoyCount benchmark, co-released by Elorian, tests whether language models genuinely count objects or rely on visual shortcuts, revealing a gap between seeing and reasoning. AI summary
Sometimes when I read a post, I get the growing feeling that I'm reading the output of an LLM. I can't necessarily define why, although there are the common tells (it's not the em-dashes, it's the load-bearing tapestries). As I do this my d…
every product i use should just be a headless mcp, that feeds into an agent, where i never have to think about anything underneath. except for our product, of course. in our product, customers love the ui and love us very much. we're specia…
TL;DR Good developers’ caution about agentic development comes from knowing that built isn’t the same as working in production. I’ve been surprised by how many excellent developers have been hesitant to go all in on coding agents. These are…
Anthropic has filed its S-1 for an IPO, but details are still scarce, including the share price. Meanwhile, Google has released Gemini 4 Carbon, a new model that tops the Artificial Analysis v5 benchmark index, but it's not yet available for public use. AI summary
Palantir CTO Shyam Sankar emphasizes the AI industry's insufficient focus on "diffusion," or the process of turning intellectual advancements into economic value. He argues that companies focused on the top line (revenue growth) are more likely to adopt AI successfully, and that the industry should prioritize workflow optimization over model improvement. Sankar also stresses the need for a systems engineering approach to AI safety, citing historical examples like Admiral Rickover's design of the US nuclear Navy as models for effective risk management. AI summary
Google's Gemini 4 Carbon, a new frontier model, has topped leaderboards in cybersecurity and long-horizon tasks, with benchmarks including 88.2% on SEC-bench Pro for vulnerability finding and 71.0% on CWE-bench for patching, with access initially limited to trusted cyber defenders. AI summary
With my kids I’m currently building a little dungeron crawler and we’re trying to learn a bit about game development along the way. However given that we live in the modern times of increasing capabilities we wanted to get a narrator and vo…
In general, LLM performance is incredibly multidimensional. There's no guarantee of a fixed ordering where the model that is best at math is also the model that's best at jokes. Despite that complexity, much of the variation in performance …
Supabase's agent-friendly features, such as local backend setup without Docker, declarative schemas, and compute alongside the database, simplify the development process for AI models and reduce friction. These improvements make it easier for developers to understand and work with the backend, aligning with the needs of agents that repeatedly create and test environments. By providing a more straightforward backend experience, Supabase enables agents to work more efficiently and effectively. AI summary
GitHub Octoverse 2026 I'm at GitHub Universe today (GitHub gave me a free ticket through their "GitHub Stars" program). EVP Jay Parikh just announced this year's Octoverse report, and highlighted these numbers in particular: <img alt="Photo…
I've written before about how November 2025 was an inflection point for coding agents - the point where Opus 4.5 and GPT-5.1 crossed a tipping point that meant their respective coding agents became reliable daily drivers for software engine…
Try Hyperagent for free, or sign in with ChatGPT to get more usage from your existing plan: https://www.hyperagent.com/login?utm_source=youtube&utm_medium=influencer&utm_campaign=FIRESHIP Are we cooked chat? Want more Fireship? 🗞️ Newslett…
Artificial Analysis v5's Intelligence Index weighs various aspects of language models, including knowledge work, coding, science, and general capabilities, but its methodology has sparked debate due to the complexity of measuring real-world usage and the trade-offs in aggregating scores. The index's emphasis on closed models may overlook the value of open models, which are still capable of performing well in specific tasks, and its shift towards private benchmarks raises concerns about transparency and accountability. AI summary
Andrej Karpathy sees a widening gap in how people see LLM capabilities. The general public has mostly not interacted with these systems. When they have, it looks like a derpy chatbot. The majority of professionals still see only a modest up…
Anthropic's S-1 filing reveals the company is profitable, with $3.9 billion in operating cash flow in H1 2026, and has raised $141 billion in capital, with $131 billion in cash and cash equivalents. The company's governance structure is complex, with the Long-Term Benefit Trust (LTBT) controlling a majority of the board but not having full authority, and the founders retaining control through their Class F shares. AI summary
If your work does not benefit from current AI tooling, that is a finding, not a failing. Document why, share it with your community of peers, bring it to your manager, and work to find solutions. An engineer who can explain where AI does no…
Meta's earnings come in the context of Zuckerberg delivering on his promise; if investors believe in Muse they must support Meta's spending on compute.
Anthropic has ended coding interviews for job applicants, citing that hunger for innovation beats tenure in hiring. This shift aims to prioritize candidates with impactful ideas and business value over technical skills demonstrated through coding challenges. AI summary
Game Boy Color on My Conference Badge I'm at GitHub Universe this week and, similar to last year , the conference badge is a full-blown color screen computer. I've been hacking around with it again, and then I ran into with Steve Francia...…
Anthropic has developed a modular architecture for their Claude platform, separating the "brain" (managed agents) from the "hands" (API), allowing for more efficient and secure model deployment. This separation enables better isolation, credentials management, and network controls, and also facilitates lower latency and better reliability. AI summary
Katelyn Lesse and Angela Jiang join Latent Space to explain the engineering behind the Claude Platform: launching Opus 5.5, separating the model’s “brain” from the agent’s “hands,” scaling APIs, and building reliable managed agents. A new i…
Developers can set up an offline coding agent using OpenCode, Ollama, and Qwen by downloading the necessary models and dependencies, setting up the project configuration, and running the coding tasks locally without internet connectivity. This setup allows for tasks such as code review, debugging, and testing to be performed offline, making it suitable for remote work or situations where internet access is limited. AI summary
Anthropic's CEO Dario Amodei defends the company's actions, arguing that most of the criticisms levied against them are driven by a negativity bias and that they have taken significant steps to address safety concerns, despite the costs. He counters criticisms of the company's approach to AI development, arguing that their "race to the top" strategy is necessary to drive progress in the field and that many of their actions, such as delaying the release of models to ensure cyberdefenses are in place, were necessary to mitigate risks. AI summary
Anthropic's IPO is finally here; the core business is very profitable, but competition may eat all of the profits, even as investors are set up for endless dilution.
Anthropic has filed for a $2 trillion IPO with a 1400% growth rate, generating 48% of revenue through cloud partners, and boasting an 85% margin, with $1 billion in operating profit in Q2 and $134 billion in cash reserves. AI summary
Get 20% off Mobbin Pro to help your agent design UIs that actually convert - https://mobbin.com/fireship Kimi K3.1 just dropped as a free 3-trillion-parameter open-weight model that comes within one point of Opus on Terminal Bench. But is i…
GPT-6.1 Terra It's always interesting to note which aspects of a new model are emphasized in the release announcement. In the case of GPT-6.1 Terra, OpenAI went with visual reasoning and computer use: We’re introducing GPT-6.1 Terra , a fas…
Because big tech platforms essentially do not have customer support and make it impossible for their users to contact a human, many people experiencing run-of-the-mill tech problems have turned to ChatGPT. And, increasingly, ChatGPT is tell…
Anthropic's draft S-1 filing omitted Q3, a section that likely discussed the company's demand performance in the third quarter, which is significant given the decline of tokenmaxxing and the rise of low-cost alternative models. This omission may indicate a lack of transparency about the company's recent performance. The missing Q3 section may raise concerns among investors and analysts. AI summary
Research: webframe: evaluated as a scraper and a sandbox webframe is a neat little project from Jeff Lindsay which lets you "run a web page like a program" - you point it at a URL or file and it executes it in a hidden headless web view and…