Executive Summary
OpenAI released GPT-5.4-Cyber, a model specialized in finding and fixing software exploits, while GPT-5.4 Pro demonstrated novel mathematical contributions. The Hermes Agent ecosystem from NousResearch continued its rapid expansion, consolidating as a leading general-purpose agentic platform. Local AI hardware demand has caused Mac Studio shortages globally, with prices in China reaching 2.3x retail. Uber reported that 11% of its live backend updates are now AI-written, signaling a major milestone in production AI adoption. Meanwhile, open-source projects are grappling with AI-generated "slop" flooding issue trackers, prompting new contribution governance models.
Key Events
- OpenAI released GPT-5.4-Cyber, a frontier model purpose-built for closed-source reverse engineering and vulnerability discovery, surpassing prior models like Mythos in exploit-finding capability. → link
- GPT-5.4 Pro made novel contributions to mathematics, with researchers comparing its impact to discovering overlooked opening lines in chess. → link
- Uber reported 11% of live backend updates are now AI-written, showing Claude Code in production-scale engineering. → link
- OpenAI Agents SDK received a major update, enabling Codex-style agentic scaling. → link
- Hermes Agent emerged as a dominant general-purpose agent, with community reaching 4,800 builders and /browser connect command enabling autonomous web interaction. → link
- Mac Studios are globally unavailable due to local AI demand, with China resale prices reaching 2.3x official pricing. → link
- Ollama launched cloud inference (claude/glm-5.1:cloud) but is capacity-constrained, actively adding GPUs. → link
- Claude Code's folder-based skills system detailed as a powerful orchestration pattern going beyond simple saved prompts. → link
- pi project introduced new contribution governance model to combat 30-50 AI-slop issues per day, with auto-close and quality-gated approval. → link
- MoE models significantly outperform dense models on unified memory systems (e.g., Gemma 4 26B-A4B vs dense 31B), key insight for Apple Silicon users. → link
Analysis
Patterns observed: The ecosystem is bifurcating between cloud-dependent frontier models (GPT-5.4-Cyber/Pro) and an increasingly sophisticated local AI movement pushing hardware to its limits. The Hermes Agent is consolidating mindshare as a multi-purpose tool, drawing users away from more bloated frameworks like OpenClaw. Agent slop in open-source maintainership has become a critical pain point, with maintainers like badlogicgames implementing structural defenses. Claude Code skills are maturing from simple prompts into composable, self-improving systems.
Escalation trends: Security is escalating on two fronts—GPT-5.4-Cyber raises the stakes for closed-source software protection, while OpenClaw's rapid security hardening through pen-testing shows the arms race in agentic tool security. Local AI hardware shortages will likely intensify as more developers seek sovereignty over their compute.
What to watch next: GPT-5.4-Cyber's real-world impact on software security practices; whether Ollama can scale cloud capacity to meet demand; the DGX Spark vs Mac Studio unified memory benchmarks; Hermes Agent's Gen 2 local agentic models; how the open-source community adapts governance to AI-generated contributions.
Tweet Feed
AI Model Releases & Research Breakthroughs
@gdb · 2026-04-15T15:25
More on GPT-5.4 Pro's latest mathematical contribution: "The closest analogy I would give would be that the main openings in chess were well-studied, but AI discovers a new opening line that had been overlooked based on human aesthetics and convention." → tweet link
@gdb · 2026-04-15T03:19
GPT-5.4 Pro for making beautiful contributions to mathematics: → tweet link
@steipete · 2026-04-15T14:33
If you look at GPT 5.4-Cyber and it's ability for closed source reverse engineering, I have bad news for you. I do very much feel the pain though, there's hundreds of teams that try to poke holes into @openclaw. Our response has been of rapid iteration and code hardening. → tweet link
@steipete · 2026-04-15T11:29
RT @PaulSolt: OpenAI shipped GPT-5.4-Cyber. A model built to find and fix software exploits. More capable than Mythos… and available today… → tweet link
@gdb · 2026-04-15T05:50
try the TurboTax app in ChatGPT: → tweet link
@Teknium · 2026-04-15T11:13
Gen 2 of local agentic models is going to be lit → tweet link
@TheAhmadOsman · 2026-04-15T14:11
There are several variables that control quality of models. These variables can be tuned to save $$$ or expand the capacity as a trade off with intelligence. Expect things to get worse as the Uber-era of subsidized tokens come to an end. → tweet link
@TheAhmadOsman · 2026-04-15T03:22
Dense models like Qwen 3.5 27B & Gemma 4 31B on unified memory are a bad idea. Simple rule: Lower memory bandwidth works best w/ fewer active parameters per token. MoE like Gemma 4 26B-A4B would work much faster on Unified Memory. → tweet link
Agent Frameworks & Developer Tools
@jezell · 2026-04-15T17:28
RT @stevendcoffey: Today, we're launching a Rather Large™ update to the OpenAI Agents SDK. Agents SDK now allows you to scale Codex-style… → tweet link
@carlvellotti · 2026-04-15T15:32
SKILLS are Claude Code's most powerful feature. But most "skills" you see are just saved prompts. The unlock: skills are FOLDERS, not files! SKILL(.)md is the orchestrator. Everything else lives next to it and loads when it's needed. [...] Reference Files, Tool Access, First-Run Setup, Self-Improving, Composition. → tweet link
@carlvellotti · 2026-04-14T18:52
The BEST and simplest way to improve AI outputs: Ask: "How can you check your work?" [...] A completely separate validation pass. AI just doesn't have the ability to reflect on its work as it does the work. You MUST have it test its work. → tweet link
@kunchenguid · 2026-04-15T02:22
where do Claude Code and Codex really differ? to get some hard quantifiable data, I benchmarked one important aspect that heavily affects our day to day usage - skill invocation. Codex worked significantly better than Claude Code at using the right skill. → tweet link
@jezell · 2026-04-15T16:55
RT @mikehostetler: I built a spec system - then trashed it. The problem isn't the tech, it's that two levels of non-determinism magnify the… → tweet link
@jezell · 2026-04-15T16:55
RT @yoonholeee: We just released code for Meta-Harness! → tweet link
@TheAhmadOsman · 2026-04-14T19:56
PRO TIP: My Agent Web Stack - SearXNG: candidate source discovery, Firecrawl: known-URL scrape and crawl, Camofox: browser fallback for JS/interaction. Search -> Extract -> Interact. → tweet link
@sudoingX · 2026-04-15T16:15
anon remember the underrated tool i tweeted about this afternoon? here it is live. i have been poking at @buildwingman from @emergent labs. [...] the self prompting is what i have been enjoying the most. it genuinely feels autonomous. → tweet link
@thdxr · 2026-04-14T18:53
in opencode 1.4.4 it no longer spawns or depend on ripgrep. we've integrated it natively into opencode thanks to @pi0's wonderful package. this is a first step before we move fully over to fff for all our search operations. → tweet link
@jezell · 2026-04-15T02:06
RT @ctatedev: Introducing wterm ("dub-term") A terminal emulator for the web → DOM rendering — not canvas → Select text, copy/paste, ⌘+F,… → tweet link
Hermes Agent Ecosystem
@sudoingX · 2026-04-15T13:22
hermes agent is becoming the general agent and the poll i dropped this morning confirmed it in public. [...] hermes agent is not a coding agent. not a research agent. not an automation agent. it is the general agent. one tool running every category of work a builder does in a day. → tweet link
@sudoingX · 2026-04-15T08:23
yo! hermes agent community just hit 4800 members. and the wild part isn't the number it's that every single one is a builder. → tweet link
@Teknium · 2026-04-15T18:27
Pliny used Hermes Agent to do the abliteration! Very Cool! → tweet link
@Teknium · 2026-04-15T17:50
Hermes on Slock coming soon!! → tweet link
@Teknium · 2026-04-15T17:58
RT @0xme66: Now Hermes can use the /browser connect command to operate the browser. I tried it out – liked a post on X. Feels quite good. Default execution policies provided. → tweet link
@Teknium · 2026-04-15T17:57
RT @elder_plinius: The crazy part? This was done (nearly) fully autonomously! Only 8 prompts from the human in the loop. Just a Hermes age… → tweet link
@Teknium · 2026-04-15T11:51
For local models, which is better in Hermes Agent? → tweet link
@Teknium · 2026-04-15T11:52
RT @nobitoshii: The new @NousResearch (Hermes Agent) dashboard looks absolutely amazing → tweet link
@sudoingX · 2026-04-15T08:14
what do you mainly use hermes agent for? → tweet link
Local AI & Hardware
@alexocheema · 2026-04-15T16:23
Local AI is the Wild West right now. Mac Studios are unavailable via the Apple website (we have a backlog of customers desperate to buy them), people are getting scammed on Reddit/eBay, and in China Macs are selling for 2.3x the official Apple price. → tweet link
@alexocheema · 2026-04-14T22:56
oMLX brought tiered kv caching to Mac. Especially important with Apple Silicon where prefill time is very long - you avoid redundant prefills, even between sessions by persisting kv caches to disk. → tweet link
@alexocheema · 2026-04-14T22:52
RT @exolabs: Apple made the M1 MacBook too good. The M1 Max was released in 2021 and has 400GB/s of memory bandwidth, even more than the NV… → tweet link
@TheAhmadOsman · 2026-04-15T01:58
Currently on my desk - 4x DGX Sparks (GB10) w/ 512GB of Unified Memory, Mac Studio M3 Ultra 512GB, Mac Mini M4 64GB. Who wants to see a comparison between the three? → tweet link
@TheAhmadOsman · 2026-04-15T04:15
Let me make local AI easy for you. Don't buy a Mac mini. Even better, mute the people that tell you that a Mac mini is good for local AI. → tweet link
@sudoingX · 2026-04-15T07:47
i dream of running sota level llms on a single 3090. and i can see it coming, day by day. [...] almost nobody has explored [kernel-level] path well. that's what keeps me awake. i'm not done with this card. → tweet link
@FrameworkPuter · 2026-04-15T15:20
RT @geerlingguy: Oops, posted early, to the benefit of EU viewers! Testing the new Arm mainboard for Framework 13: → tweet link
@TheAhmadOsman · 2026-04-14T22:15
What am I working on? Condensing everything I do into one place: local AI / LLMs, inference + benchmarking, hardware + cluster builds, LLM research + notes, agent workflows, real-world perf. All into a single, searchable, indexable platform. > The Buy a GPU Website. COMING SOON. → tweet link
@TheAhmadOsman · 2026-04-15T00:20
Alternatives to the BLOATED OpenClaw? - Hermes Agent - ZeroClaw - Pi (OpenClaw is built on top of it) - NanoClaw. Recently I also came across GitClaw which has an interesting design. → tweet link
@TheAhmadOsman · 2026-04-15T00:02
In a future where tokens quantity and quality will determine your standing and wealth, fighting for compute sovereignty is a worthy battle. → tweet link
Open Source AI & Infrastructure
@ollama · 2026-04-15T04:46
ollama launch claude --model glm-5.1:cloud (We are rushing to get more capacity 🙏🙏🙏) → tweet link
@ollama · 2026-04-15T03:58
We are rushing to add more capacity to Ollama's cloud. Please be patient with us as we add more GPUs. 🙏 → tweet link
@victormustar · 2026-04-15T08:44
Open source AI music is actually good now 🔥 Made a free demo for ACE-Step 1.5: describe any song, get it back in seconds ⤵️ → tweet link
@victormustar · 2026-04-15T07:49
RT @RyanLeeMiniMax: No.1 Again! 🎉 MiniMax M2.7 has only 230B parameters with 10B activated, yet delivers… → tweet link
@victormustar · 2026-04-14T20:47
RT @RyanLeeMiniMax: I just updated our license. For personal use, you're free to run the software on your own servers for coding, build… → tweet link
@jezell · 2026-04-14T20:09
RT @firecrawl: Introducing Fire-PDF, our new Rust-based parsing engine 🔥 - Convert PDFs into markdown 5x faster - Extract full tables and… → tweet link
@ollama · 2026-04-14T18:59
RT @lmsysorg: Excited to be part of @ollama Gemma Day tomorrow in Palo Alto! Khoa Pham (@kwafam7) from @radixark will demo Gemma 4 in produ… → tweet link
@ollama · 2026-04-14T18:59
Ollama and Google Gemma team is hosting an Ollama Gemma Day in Palo Alto tomorrow night (Wednesday, April 15th at 6pm). Lots of amazing speakers from @GoogleDeepMind and @sgl_project / @radixark! → tweet link
@tinygrad · 2026-04-15T09:02
We're hiring from the pool of tinygrad contributors. Hybrid in-person/remote, offices in San Diego and Hong Kong. In the era of slop, come help build something beautiful. → tweet link
@badlogicgames · 2026-04-15T15:26
RT @ClementDelangue: Weird how some people always target open-source in AI! First it was: "Open-source AI will destroy the world" (spoile… → tweet link
Software Engineering & DevOps
@badlogicgames · 2026-04-14T21:39
People of pi. The great @steipete has graced our repository with a bespoke slop PR to fix cache affinity in the OpenAI Responses provider. [...] And the new "pi contribution model (tm)" is now live: auto-close issues/PRs, quality-gated approval, "lgtm" for well-written contributions, blocked accounts for agent slop. → tweet link
@hnasr · 2026-04-15T14:36
I wrote a new book called Root Cause, for those who enjoy the art of backend engineering. [...] 15 chapters, each a story about a backend bug, with investigation, diagrams, a section of a fundamental concept until the root cause is revealed. → tweet link
@MatejKnopp · 2026-04-14T22:15
Today Codex wrote 190 lines of code for me, 100 of which were completely unnecessary. Took me about same time prompting to get rid of the shit as writing the thing myself. But then again, it also wrote a decent test. → tweet link
@thdxr · 2026-04-15T17:04
we've been in deep refactor / cleanup mode. on one hand opencode has been super useful… but i also wasted an hour today trying to port a single module to a new structure - opus, gpt, kimi all failed. → tweet link
@steipete · 2026-04-15T18:27
That was the case in December. 4 months and thousands of work hours later, we have a great security concept; you can go all yolo, use a sandbox (Docker or OpenShell), there are allow-lists and per-access exec allow/deny prompts. There's hundreds of security researchers that pen-tested it. → tweet link
@badlogicgames · 2026-04-15T13:22
github tries to be a single page app and sucks terribly at it. i want my full reloads back... → tweet link
@jezell · 2026-04-15T10:46
RT @bibryam: Your Container Is Not a Sandbox → tweet link
@nummanali · 2026-04-15T11:05
I tried the new Warp update. It's nice but too opinionated for me. Lacks workspaces, and tabs in a pane. cmux is too good: Workspace → Pane → Surface → Panel. → tweet link
@MengTo · 2026-04-15T15:15
So this is a highly curated collection of 400+ DESIGN.md that's editable and promptable. Just one click to change styles, turn into branding, mobile versions, slide decks and site sections. → tweet link
Industry & Business
@jezell · 2026-04-14T22:02
RT @wallstengine: > Be Uber > push Claude Code across engineering > 11% of real, live backend updates now written by AI > use AI for rid… → tweet link
@TrungTPhan · 2026-04-15T18:02
RT @bearlyai: Jensen on how Nvidia prioritizes chip sales. He says it is "first in, first out". Company has to make a purchase order, but… → tweet link
@TrungTPhan · 2026-04-15T17:14
RT @bearlyai: Credo Technology has been interesting AI supply chain play: Stock is up 13x in the past 5 years to a market cap of $29B… → tweet link
@jezell · 2026-04-14T21:28
RT @RihardJarc: The CPU shortage is severe. The comment from $AMZN regarding the current demand for their Graviton CPUs is not getting enough attention. → tweet link
@jezell · 2026-04-14T22:10
What's cheaper, buying licenses for all your agents, or using them to vibe code a Word replacement? Microsoft is about to find out the answer. → tweet link
@jezell · 2026-04-15T02:05
This was inevitable. Supply and demand. If you make it so expensive no one can afford it, your capacity problems go away. → tweet link
@gospaceport · 2026-04-15T15:06
Even the OSS-with-a-side-of-SaaS market is scared at what is coming down the pipeline, massive competition from all angles. → tweet link
@LinusEkenstam · 2026-04-14T20:29
[Higgsfield Marketing Studio] Paste a product link → get 9 different video ad angles (UGC, unboxing, review, tutorial, TV spot). Powered by Seedance 2.0. $0.347 per generation. → tweet link
Flutter & Cross-Platform Development
@jezell · 2026-04-15T17:57
I don't get the "Flutter Web is a waste of the Flutter team's time" takes. If all you want is single codebase mobile apps, you should be using React Native. Literally the only nice thing about Flutter is that it includes web. → tweet link
@jezell · 2026-04-15T02:41
All of these problems are just as solvable in Flutter Web as they are in Figma and Google Docs... but the defaults definitely aren't right and some basic things like this shouldn't be so hard to enable. → tweet link
@MatejKnopp · 2026-04-15T08:57
Flutter web has some warts, but this ain't it chief. It's like complaining that you can't select UI in Discord or Slack. Btw. Ctrl+f doesn't work in Github Pull requests code any more. → tweet link
@ASalvadorini · 2026-04-15T12:17
RT @ulusoyapps: Flutter Web has its place in the ecosystem. It may not work perfectly in all use cases, but at Wolt its business impact is… → tweet link
@ASalvadorini · 2026-04-15T05:31
Once again, you shouldn't use GetX and you should be aware of the build context, this applies to both Android and Flutter as they've the same architectural design. → tweet link