Aidan Pak

It turns out, to solve the integration problem, all we needed to do was give the agent access to bash + filesystem + browser, and it could integrate with effectively anything. One of the biggest takeaways from OpenClaw (and Claude Code/Codex) is that SaaS companies have already lost the battle for the integration surface, whether they've realized it or not. The traditional driver of growth/expansion for enterprise software was owning the platform where real work is done (system of work). While the moat is rooted in being the system of record, owning the interaction layer on top is what justified seat expansion, feature upsell, and pricing power. The future value of enterprise SaaS companies like Salesforce is entirely dependent on their ability to maintain this status as the system of work, and in the age of AI, that means owning the agent, or the interface with the agent. But what Claude Code / OpenClaw are showing us is that if you give the agent bash, a filesystem, and a browser, the agent can just "learn" to use your product and harden it in a "skill." This could take the form of a script calling an API, using a CLI, or using the browser to navigate the UI. This doesn't mean that systems of record get fully replaced in the medium term, but it does mean they're beginning to lose their status as the system of work (essentially becoming just a data provider) which significantly questions future growth and pricing power.

Generative media is one of the most exciting spaces to follow. Melius is making AI delightful for everyone.

Every category of knowledge work desperately needs a platform built for human/agent collaboration, and Melius has nailed that product surface for creative work. Through the Melius API, CLI, and MCP, agents get every frontier image, video, and audio model behind a single endpoint, and a node-based canvas that gives creative work a structure they can actually operate on. Creatives get a beautiful canvas for collaborating with those agents, and the right surface to express judgment and taste. Congrats @n0w00j and team!

Creative tools like Higgsfield require excessive prompting and have predatory pricing, and fail to keep generations consistent. Even worse, you’re forced to go back and forth on different pages and stitch together assets on different platforms.

Looking at the model card, the difference between Fable and Opus 4.8 looks small. Same for GPT-5.4 vs 5.5. But after using Fable for a while now, the thing that really stands out is its ability to reason for longer without falling apart and the capability gains from duration is massive. The majority of agent tasks are compositional: a task is mostly a chain of dependent steps, meaning it only finishes if ~every step lands. So task success ≈ p^t (where p = per-step accuracy, t = number of steps). For example, the jump from 98% to 99% roughly doubles the length of task completable with 50% probability. This is why a single number on a system card is the wrong lens to evaluate model capability and instead performance should be measured as a function of compute and the length of time it can stay coherent.

Very well said and I would add there's a pareto distribution of task value. Open models can already do 80-90% of tasks and will continue to over time (especially now with post-trained OSS >= closed-source models for a specialized task), but even as OSS models improve, 80-90% of spend will continue to go to the frontier because the economic value is wildly skewed toward the model that can solve the hardest 5-10% of societal problems.

What's next? Haiku Sonnet Opus Fable (guardrailed Mythos) Mythos

@AnthropicAI Cant wait to use Fable to "grammatically correct this email" 💀

Claude Fable 5 will be available again globally tomorrow. After a series of productive conversations with the US government, we're redeploying the model with a new set of classifiers to target and block more cybersecurity tasks. In the near term, some routine tasks like coding and debugging will fall back to Opus 4.8. We’ll continue to refine these classifiers over the coming weeks to reduce false positives and better distinguish genuine misuse from legitimate requests. We’ve also begun drafting a consensus framework—with Amazon, Microsoft, Google, and other Glasswing partners—for assessing the severity of AI jailbreaks and how AI developers should respond to them. We invite other industry partners and model providers to join us in this effort. Finally, we’re scaling up our collaboration with the US government on model testing and safeguards. This will include pre-release access to models and safeguards for evaluation, information sharing on jailbreaks and misuse, and dedicated resources for joint research. Thank you to our users for your patience, and to our partners across the government, industry, and the research community who worked alongside us to make Fable 5 available again. Read our full blog: anthropic.com/news/redeployi…

@databricks is uniquely positioned to be the first to actually solve HTAP, the holy grail DBs have chased for decades. HTAP has historically failed because no single layout serves both workloads. OLTP wants data stored by row to grab one record fast, OLAP wants it by column to scan a field efficiently, and you cannot be optimal at both. So in practice you run two systems, an OLTP db and an analytics warehouse, glued together with CDC, which is incredibly brittle and leaves you fighting replication lag, duplicated spend, and split governance. Following the @neondatabase and @mooncakelabs acqs, Databricks' bet is to stop trying to unify the engine and unify the storage instead. Neon (now Lakebase) is OLTP that already split storage from compute and writes Postgres straight to object storage. Mooncake (pg_mooncake) is the transcode piece that keeps a columnar Parquet/Iceberg mirror of those tables continuously in sync, so Postgres keeps serving transactions while engines like Spark and Photon run analytics on the same fresh copy, all governed by Unity Catalog under one permission and lineage model. Of course there are still tradeoffs and this won’t be true real-time HTAP, but killing the pipeline and unifying storage seems to be the right direction, and if anyone pulls it off, it is Databricks.

@0xhappier @JahowYu @trymelius Not on the visvim.... @JahowYu

@evisdrenova This is the case for intelligent routing between closed and oss models. @OpenRouter is great and I also think @databricks has a strong case to own routing for large enterprise

I'm starting to believe that every company will run open source models in addition to OAI/ANT. I think we'll see more OSS models trained in the US as well. With that, the demand for infrastructure to support the open source models will dramatically increase.

look at the melius fellas man

It's fascinating to see how frontier capabilities seem to be first unlocked via the harness, but over time, that behavior gets inevitably baked into the weights of the model and the scaffolding around it evolves. For example, in the early days, agent frameworks wrapped LLMs in pre-defined DAGs with dedicated planning, context management, and tool-calling steps to overcome limitations in the base model such as hallucinating tool arguments, forgetting instructions, and context anxiety. These encoded assumptions quickly became stale as reasoning models were introduced and quickly improved, but importantly, that didn't mean the infrastructure around the model went away. What's super interesting about SPIRAL is that it's research on how to encode the current SOTA behavior of a harness (e.g. Claude Code dynamic workflows / loops) directly into the weights. Specifically, training a model to spin up parallel reasoning traces and then synthesize them into a final answer (which is one of the multi-agent architectures that is prevalent with dynamic workflows, e.g. fan-out-and-synthesize) but instilled end to end via RL rather than in the harness. I have been incredibly impressed by the capabilities unlocked by dynamics workflows. This is where Claude writes the orchestration as a JS program and then runs that code. The script holds the loop, the branches, and the intermediate results, fans out a pile of subagents in parallel, and a synthesizer folds them back into one answer. Other multi-agent patterns like tournaments, adversarial verification, and generate-and-filter are also prevalent. This behavior is currently a part of the harness but might soon be embedded in the weights.

@StockSavvyShay RIP consumer electronics

$MU ABSOLUTELY CRUSHED THEIR EARNINGS • Revenue $41.5B vs Est. $35.5B • EPS $25.11 vs Est. $20.39 • Net Income $33.7B vs Est. $23.9B • Gross Margin 85% vs Est. 82% Q4 Guide • Revenue $50B vs Est. $43B • EPS $31.00 vs Est. $25.07 • Gross Margin 85% vs Est. 84%

Great post! I think the next step for "code review" looks less like “review” and more like code validation / dynamic verification. It will be less statically reading a diff / running tests, and more simulation where agents drive an environment, run a feature against its real dependencies, check how it behaves against other services, and adversarially try to break it. The problem today is that productizing this is incredibly hard: (1) dev environments are custom to each enterprise, and (2) the cost of compute to run them at scale is still too high. I believe @cognition is heading in this direction

@JakeEpstein1 Super cool stuff Jake! Offloading as much work as possible to small, local OSS models will be increasingly important. Would be great to add in a smart router / BYO keys to route between frontier / local models

I built Bayou, an open source harness for running OSS models on Mac. It'll run models larger than available memory (quantization + MoE caching) - even spec decode against a small model on an iPhone. Building is how I learn to invest. Run it yourself: jakee.vc/bayou

@leylndd @GeoffreyHuntley VCs when Sequoia leads the A

“independent thinkers” as soon as boris chreny breathes

Longer-running agents are creating enormous KV cache that has to be moved around: offloaded across storage tiers, shared between nodes, etc. The reason CPU demand is exploding isn't that CPUs are replacing GPUs (Matmul still belongs on the GPU) which is purpose-built for it. It's that CPUs are increasingly the ones moving all that KV and data around to keep the GPUs fed along with the actual computing required to execute tool calls, navigate a browser, etc.. It's why NVIDIA's NVL72 rack pairs the GPUs with Grace CPUs.

An NVIDIA executive just told me the GPU-to-CPU ratio will go from 2:1 today to 1:1 “in months" due to agentic AI. NVIDIA executive is saying 1:1 IN MONTHS. CPU CPU CPU $NVDA

@n0w00j 🤣

$100m ARR and can't even spend $11 dollars on claude tokens to build a simple cancel plan flow. "send us a note" LMAO yea okay

@Jason I've heard from a source close to a Mag 7 CEO that the government is quietly pressuring large companies (Microsoft, Meta, etc.) to avoid mass layoffs. Many leaders obviously see an incredible opportunity to optimize their headcount given AI but being forced to slow it down

There will be significant AI job displacement. The only question is how fast we can manifest new jobs. The AI washing claim was accurate for the first couple of rounds of layoffs, but the ones this year are explicitly because owners think they can do more with far less. And the owners are correct. You can easily job proof yourself by being AI-first, starting a company or joinging generation toolbelt.