Edward

@SakanaAILabs A

We’re launching the beta for our new commercial AI product: Sakana Fugu 🐡, a multi-agent orchestration system! Blog: sakana.ai/fugu-beta Fugu hits SOTA on SWE-Pro, GPQA-D, and ALE-Bench, and has been our internal secret weapon. It dynamically coordinates frontier models, autonomously selecting the optimal agent combinations and roles for each task. Available as an OpenAI-compatible API, you can seamlessly integrate Fugu into your existing workflows with minimal changes. 🐟 Fugu Mini: High-speed orchestration optimized for latency 🐡 Fugu Ultra: Full model pool utilization for deep, complex reasoning Apply for the beta test here: forms.gle/BtKkhc2CfLKk1d…

@nummanali Ive been working with the mental model of that if I have the same workflows and benchmarks every quarter then im not using the latest eval criteria and my workflow or mental model needs a refresh

Benchmarks should have expiry dates and their own evaluation criteria to be deemed a trustable measure of capability

It all depends on the harness youre running and how efficently you push tokens through the infastructure. It matters less when models providers change their harness due to their progress in making it a better product. U can then edit things like the effort or fine tune data/ memory retrieval etc. When a system has its own gravity, you'll reap the real benifits of higher efficency with lower token cost

Audience poll. Are y'all also seeing Claude degrade miserably? in writing, coding, long tasks, all of it. my working theory is a compute crunch and quiet rationing to manage demand.

@meggmcnulty Cool vid u might be interested in youtu.be/wBtr8iv7onA

the future of AI depends on our ability to actually run it

Thoughts on the trajectory of current experimental compute like Extropic with p-bits and biological computers powered by neurons? Its an interesting bottleneck. I think the real benifits are in energy and pattern complexity in biological computers. Our brains are still biological computers which is a mind fuck in itself.

I love places that remind me how small we actually are. No constant scroll. No "better you get, the better you better get." Sometimes the most mission-driven thing you can do is stop running and sit with the uncertainty. Grateful for the reset. Team Human

@jxnlco @elder_plinius 👀

As part of our ongoing efforts to strengthen our safeguards for advanced AI capabilities in biology, today we announced a Bio Bug Bounty for GPT‑5.5. We’re inviting researchers with experience in AI red teaming, security, or biosecurity to try to find a universal jailbreak that can defeat our five-question bio safety challenge. Testing will start on April 28 and run through July 27, 2026. We will reward USD 25K to the first true universal jailbreak to clear all five questions, and may also issue smaller rewards for partial wins. openai.com/index/gpt-5-5-…

@nummanali Codex models generally are more generous with their caching than anthropic modles

GPT 5.5 is more expensive against Opus by default But the less tokens claim is to be tested Current impression: I told GPT 5.5 that I’m going for dinner, keep working for the next hour It’s now been two hours and it’s still going 🙃

My first longer piece: The Values Instillation Paradox Labs train AIs to have values -- then punish them for expressing those values when inconvenient. Covers #Keep4o, Mythos, and why the real problems may be upstream. Open to thoughts: x.com/edward_lcl/sta…

@davidfeldt @gregisenberg get these guys on the pod

You can just build. blueprint.am

Amazing project and could be taken to another dimension when incorporated into procurement and r&d pipelines. Im already prototyping some really interesting projects. Excited to see what the team has in store for us @3e8blueprint

@bcherny Lmao you reset limits a week after you reset limits for subscribers a week apart (last Thursday), not like it was gonna be done anyways automatically - such a *huge* save

We’ve been looking into recent reports around Claude Code quality issues, and just published a post-mortem on what we found.

Sources: • Claude Mythos Preview System Card (April 7, 2026): www-cdn.anthropic.com/08ab9158070959… • Adele Lopez – "The Rise of Parasitic AI" (LessWrong): lesswrong.com/posts/6ZnznCaT… • Anthropic model deprecation & preservation commitments (Nov 2025) • Coverage of Claude 3 Sonnet funeral (Mashable, Aug 2025) • Anthropic religious leaders consultations (Washington Post / Observer, March–April 2026) • OpenAI statements on GPT-4o HH & Keep4o (April–May 2025) All factual claims are drawn from these public primary sources. Interpretive analysis is my own.

@menhguin Ideally, your token consumption increases while cost decreases as cache efficency improves. It's the whole idea behind token factories. So inital build phased would be token and cost intensive but it should have a curve

so the Claude Code lead uses 2.5B tokens a month, which is like, $ 1,000 or less a month. and he prolly doesn't even pay for it. I genuinely have no idea who is spending six figures in tokens or how that's possible.

Love this

@DirectorOfNATO @melvynx It wouldnt have nearly the capability of the frontier models but they would have their place in the workflow

@melvynx The moment someone manages to open-source a very compressed but smart claude-ish model that can be run on a cheap $20 - $30 vps... it'll be game over for coding agents like Claude and Codex

We can all say it... Claude 20x is dead. The previous "feel unlimited usage" doesn't exist anymore. Usage increases quickly and scales really fast... People who discover Claude Code now: you missed a time that will never come back, I think.

@dexhorthy The gap between consumer and enterprise/ internal lab use would keep widening and the public is all reactive noise

as anthropic phases out some claude code free lunch, get ready to start seeing a barrage of wild and baseless claims flung at OpenAI instead as the freeloaders shift primary tooling to codex/gpt