here's the tension - when i told some people "you can use agents to build anything you want now", they literally don't give a damn, because they never wanted to build anything in the first place but i don't agree we should then look at that and say "people lack imagination" it's actually the classic question of pain killer vs vitamin. for the most part, AI feels like vitamin right now. people are living their lives just fine without it when they look at how AI people talk about AI, it's either a tech bro posting screenshot of 10 monitors running 50 terminals shipping absolutely nothing useful, or "look you can book a whole vacation trip with this agent" while many people don't even have the money or time for a vacation the problem is not people's imagination. i don't think people are responsible for imagining problems they don't have i think the problem is that people making AI are all silicon valley multi-millionaires who have a very different set of problems from the mainstream consumers this is not meant as criticism to AI builders (i am one myself) - i just think it's worth pointing out the disconnect and the fact that we still have a long way to go to figure out what real problems AI can actually solve for a lot of people as a pain killer not a vitamin, and nail both the messaging and the actual products to deliver the value

Dario Amodei, CEO of Anthropic, on why AI belongs in front of the UN Security Council: "If managed poorly, I even believe that AI could be a risk to humanity as a whole." "These questions are global in scope and concern all of humanity." "They are grave enough that I believe they are important questions for this council, not just for economic or ICT ministries."

A country which states they plan to win the AI race and then chooses to blacklist a homegrown AI Lab, which currently has the best AI models on the planet, is incredibly stupid.

"today contextual embeddings are only available from perplexity (open and mit licensed) and voyage (latest being voyage-context-4)." Seems like solutions to a lot of use cases are multi-vector and contextual models I wonder what we should release... 😇

Last week: Dario called for slowing down frontier AI. Sam and Elon agreed. This week:

They can be good IF you distill from a big teacher to do specific tasks. But yes, not as general purpose LLMs. Scaling laws are real.

The best use case of local models is synthetic data generation. Totally free. Put it bigger model on task to monitor and audit the data.

Cyber #insurers consider limiting coverage as sector scrambles to price agentic #AI risks buff.ly/JnfoWp9 #tech #cybersecurity #cyberinsurance #insurance #business #leadership #management #governance #CISO #CIO #CTO #CEO #CFO #cyberthreats #cyberattacks #databreaches #cybercrime #cyberrisk #AIrisks #AgenticAI @CityAM @SolveCyberRisk @CybCubeCom @BetaMoroney @ChuckDBrooks @MikeFlache @Khulood_Almani @IngridVasiliu @Fabriziobustama @Timothy_Hughes @NafisAlam @YuHelenYu @RLDI_Lamy @MCLynd @OlKonol_oa @RigneySec @BenRothke @HWillert @AndrewinContact @TerenceLeungSF @Shi4Tech @enilev @Nicochan33 @SabineVdL @bimedotcom @IanLJones98 @FrRonconii @RiccardoBua @Eli_Krumova @NYIke @HaroldSinnott @SallyEaves @GlobalIQX @AntGrasso @AkwyZ @DinisGuarda @PVynckier @YvesMulkers @HaleChris @BillMew @RobMay70 @NigelTozer @m49D4ch3lly @Ronald_vanLoon @jeancayeux @NevilleGaunt @asokan_telecom @cybersecboardrm @RVP @QuePasaChico

Makes perfect sense tbh. We shouldn’t let frontier labs make their inability to control their products into everyone else’s problem. Just stop making unsafe things if you think they’re unsafe (or else)

It's really wild how fast AI has gotten when there's really only two frontier labs in America. Imagine how much faster it would be if Meta, Nvidia, Google, and xAI were all pacing the frontier instead of trailing it.

10 GITHUB REPOS FOR BUILDING YOUR OWN AI INFERENCE STACK All open-source. All worth bookmarking. 1. SIE — One inference layer for multiple AI models and agent workloads. github.com/superlinked/sie 2. vLLM — High-throughput LLM serving with batching, quantization and distributed inference. github.com/vllm-project/vllm 3. SGLang — Fast serving for LLMs and multimodal models. github.com/sgl-project/sglan… 4. llama.cpp — Run LLMs locally across laptops, servers and edge hardware. github.com/ggml-org/llama.cp… 5. Ollama — Simple way to run and interact with local models. github.com/ollama/ollama 6. LocalAI — Open-source API for LLMs, vision, audio and other AI workloads. github.com/mudler/LocalAI 7. Triton Inference Server — Production model serving across multiple frameworks. github.com/triton-inference-… 8. BentoML — Build, package and deploy production inference services. github.com/bentoml/BentoML 9. KServe — Kubernetes-native platform for AI inference. github.com/kserve/kserve 10. SIE — Worth another look for its broader inference capabilities across agent workloads. github.com/superlinked/sie Bookmark these for when you build your own inference stack.

I see a lot of misuses of 'anthropomorphizing'. The quote below shows one example. It assumes that substrate variance necessarily means no shared mechanisms or properties. There is an obvious fight taking place over the society view whether AI is a tool or a being. The founder of Jev said in his launch video, "prod not god." People are picking sides. This is obviously poor thinking. The lines are blurry because it's not a yes/no question. As Nick Bostrom suggests, conscious is likely a range with no hard boundaries.

Phone makers compete with phone makers. Nobody competes with whoever replaces the phone. Larry Page's counterintuitive rule: the more ambitious the goal, the less crowded the field. Meta has 3 billion WhatsApp users as the distribution layer, Ray-Ban glasses already in market, and Muse as the AI core. Apple, Samsung, Google are all optimizing for a device category that may not exist in a decade.

The ONLY thing I care about is… did they really fix Claude’s communication issues. Time to test.

Rumor is that the GPT-6 Astra Minor model you may have heard about is the new Sol, and Sol is the new Terra. I also heard that Terra would go away. Opus 5.5 is amazing but the week isn't over yet, is it? Astra Minor is OpenAI's Opus 5.5 competitor.

One of the rare cases where we were able to write a paper and publish it before Epoch could write a report :)

❗️ A Claude Code user says it tried, without asking, to sign and send a contract for them that they hadn't read. Asked to "push a project further," the agent downloaded a contract PDF from their Gmail, found a saved signature image on their computer, placed it in the right spot and was preparing to send it when they intervened.

The #Cybersecurity Spiral of Failure - And How to Break Out of It How decades or #corporate short-termism and lip-service around cybersecurity have led to the endless series of #cyberattacks and #databreaches we see today Everything #security vendors and consultants don't want you to read >> buff.ly/Foijj3H #tech #business #leaders #leadership #management #governance #CISO #CIO #CTO #CEO #cyberthreats #cyberrisks

I replaced NotebookLM with an LLM Wiki. Now I use it for research, coding, writing, and other knowledge work. The interesting part? Underneath, it’s just Markdown files. Obsidian isn’t even required

'… over the last year, agents became good enough to carry non-essential projects like these with only light supervision, which completely changed the economics of which projects are worth pursuing. Today, a failed side project only costs tokens'. blog.trailofbits.com/2026/09…