Executive Summary
The past 24 hours in tech and AI have been dominated by the controversial release of Claude Opus 4.7, with widespread user reports characterizing it as a regression from 4.6—slower, more expensive, and less reliable—potentially fueling a migration to competing tools. OpenCode 1.4.11 emerged as a notable challenger, introducing beta workspace support for git worktrees and remote environments, with the platform rapidly evolving into an agentic IDE. Ollama 0.21 shipped native support for Nous Research's Hermes Agent, alongside a $25k creative hackathon, signaling increased momentum for open-source agent frameworks. In the local inference space, quantization techniques advanced significantly (GLM-5.1 compressed ~75%), and RTX 3090 continued to be praised as the best-value GPU for home AI setups.
Key Events
- Claude Opus 4.7 faces widespread criticism as regression. Multiple developers and benchmarks report worse performance than 4.6 despite 50% higher token costs, faster rate limits, and increased hallucination. → tweet
- OpenCode 1.4.11 adds workspace beta support for git worktrees and remote environments. → tweet
- Ollama 0.21 ships with native Hermes Agent integration, enabling one-command local deployment of the self-improving agent. → tweet
- GLM-5.1 quantization tutorial: PRISM-Dynamic Quant achieves ~75% compression (747B → ~385GB) with tensor-level precision tuning. → tweet
- Hermes Agent Creative Hackathon launches—16 days, $25k in prizes, presented by Kimi MoonShot and Nous Research. → tweet
- New Kimi paper on inference optimization: Lower latency, higher throughput, cheaper tokens. → tweet
- RTX 3090 maintains "GOAT GPU" status for local inference value in 2026, with analysts warning GPU prices rising due to HBM shortage. → tweet
- Codex evolves toward full agentic IDE: Proactive suggestions, Session Graph plugin, sound notifications via Pulse Audio. → tweet
- Claude Design launched with limited availability, generating buzz about Anthropic's expansion beyond coding. → tweet
- Local DNA sequencing with AI: Developer runs Evo 2 (40B-parameter DNA LLM) on DGX Sparks. → tweet
Analysis
Claude vs. Codex Dynamics: Opus 4.7 criticism is significant—the developer community is actively comparing it to Codex, with some declaring switching to OpenAI's tool. This could mark a turning point in the coding assistant market. Codex's rapid feature development (workspace, plugins, otel tracing) contrasts with Anthropic's perceived misstep.
Agent Framework Proliferation: Hermes Agent's Ollama integration and the associated hackathon signal strong momentum for open-source agent frameworks. The model-agnostic approach (works with Claude, GPT, local models) is gaining traction.
Local AI Maturation: Quantization advances (GLM-5.1 to 385GB), multi-Mac inference clusters, and voice dictation integration into coding workflows indicate local AI is moving beyond experimentation toward practical daily-driver use cases.
Hardware Trends: RTX 3090 continues to dominate local inference recommendations despite newer GPUs, highlighting the HBM shortage's impact on accessibility. Analysts warn a price window for consumer GPUs is closing.
What to Watch: Continued Opus 4.7 vs. Codex user migration, hackathon submissions showcasing Hermes Agent capabilities, and whether Anthropic responds to the criticism with fixes or acknowledgment.
Tweet Feed
Claude Opus 4.7 Controversy
@RydMike · 2026-04-17T21:26
RT @argofowl: opus 4.7 is disappointing as i predicted anthropic is lost they think their users are dumb and will eat it all up → tweet
@RydMike · 2026-04-17T21:23
RT @petergostev: BullshitBench: Opus 4.7 did WORSE than Opus 4.6 family. The 'Max' thinking version did worse than non-thinking - 74% 'push… → tweet
@RydMike · 2026-04-17T21:21
RT @Akasheth_: unpopular opinion: 100$ in codex > 200$ in claude → tweet
@nummanali · 2026-04-17T20:52
Opus 4.7 isn't showing thinking summaries in Claude Code Solution: claude --thinking-display summarized → tweet
@victormustar · 2026-04-18T09:50
RT @0xSero: Opus-4.7 is unusable. Multiple times i have given it specific links, for it to use, specifically. Instead it goes finds unrel… → tweet
OpenCode / Agentic IDE
@thdxr · 2026-04-18T17:15
RT @jlongster: Excited to release workspace support (beta) in OpenCode 1.4.11: lets you run sessions in git worktrees and even remote envir… → tweet
@thdxr · 2026-04-18T16:18
i had opencode write me a plugin that lets it play sound over pulse audio server when a session is idle this means it works even though it's running on a remote machine linux audio is good actually → tweet
@thdxr · 2026-04-17T19:32
we've done instrumentation work in opencode recently so we can get traces out they get shipped to a local otel tui (by kit) which also has endpoints for the agent to read this data provides a good feedback loop for the agent to analyze issues - will do a video soon → tweet
@gdb · 2026-04-17T19:46
codex for proactively suggesting what it can do for you: → tweet
@gdb · 2026-04-18T05:34
codex is becoming a full agentic IDE → tweet
@gdb · 2026-04-18T09:52
codex makes work plain fun → tweet
@thdxr · 2026-04-17T19:52
the biggest impact on my coding workflow lately hasn't been anything agent related it's realizing how good these parakeet local voice models are and then just dictating everything into opencode typing is more of a burden/bottleneck than you realize → tweet
Hermes Agent & Ollama
@ollama · 2026-04-17T23:26
ollama launch hermes Ollama 0.21 includes supports Hermes Agent, the self-improving AI agent built by @NousResearch. → tweet
@ollama · 2026-04-18T00:55
RT @NousResearch: Hermes Agent 🤝 Ollama → tweet
@badlogicgames · 2026-04-18T16:12
RT @matteocollina: Regina is a production-ready agent orchestration layer built on Platformatic Watt. You define agents in Markdown. Start… → tweet
@crystalsssup · 2026-04-18T04:33
RT @NousResearch: The Hermes Agent Creative Hackathon starts now 16 Days, $25k in Prizes Presented by @Kimi_Moonshot & @NousResearch → tweet
@Teknium · 2026-04-18T01:04
Happy to hear it! → tweet
@Teknium · 2026-04-18T09:04
RT @GitTrend0x: Hermes 一丢 Agent，全网又卷出 5 个新进化体！ Nous Research 的 hermes-agent（96k+ stars）底层太能打了：持久记忆 + 自动提炼技能 + 跨会话成长，社区直接当 DNA 疯狂 remix。 → tweet
@Teknium · 2026-04-18T08:31
RT @Kimi_Moonshot: We don't know what Kimi + Hermes agents will look like in the wild yet. That's why we built the hackathon. Presented by… → tweet
@sudoingX · 2026-04-18T01:10
RT @sudoingX: people ask me this all the time so let me say it out loud. the best model to pair with hermes agent is claude opus 4.7, by fa… → tweet
Model Releases & Research
@crystalsssup · 2026-04-18T12:14
Fresh paper from the Kimi team! TL;DR: Lower latency, higher throughput, and cheaper tokens. → tweet
@Ex0byt · 2026-04-18T16:02
Sharing for anyone curious about quantizing GLM-5.1 without flying blindly. Out of curiosity, I ran a tensor-by-tensor sensitivity eval with PRISM-Dynamic Quant to find where the model can be compressed without meaningful degradation. PRISM-DQ Pareto analysis on GLM-5.1 (747B params, 78 blocks): • Knee: 4.12 BPB (point before quantization errors start compounding) from a Q5_K base • 63 tensor-level overrides across 32 blocks (per-block adjustments based on sensitivity) • Most sensitive, kept at higher precision: token_embd (0.854), output (0.853), ffn_gate (0.805) • Least sensitive, quantized more aggressively: attn_output (0.432), ffn_up_exps (0.446) Expected Size Result: ~1.5 TB >>> ~385 GB. Recipe below. A REAP pass (expert pruning) on top can reduce it by another ~30% safely. → tweet
@TheAhmadOsman · 2026-04-17T20:32
RT @TheAhmadOsman: Currently running GLM-5.1 locally Cannot believe this thing is running on my own GPUs, its really smart → tweet
@alexocheema · 2026-04-18T15:21
RT @corbett3000: It's Friday night. I'm in SF. I'm coding applications on a local mac mini + M2 @exolabs cluster running Qwen3.6. This is… → tweet
@alexocheema · 2026-04-18T15:16
RT @izzycodev: Running Qwen 3.5 at ~30 tok/s locally across two Macs. M4 Mini + M1 MacBook Max using @exolabs → tweet
@alexocheema · 2026-04-18T00:12
people are now sequencing their DNA at home, locally on DGX Sparks and Mac Studios. this madlad is running Evo 2, a 40B‑parameter DNA LLM that predicts genome sequences instead of text. local AI is going to unlock a world of creativity. @karpathy's personal computing v2 is here. → tweet
Local Inference & Hardware
@sudoingX · 2026-04-18T10:30
no prompt engineering, no agentic harness, no tool calls. just me being lazy in llama.cpp's web ui and gemma 4 31b dense taking the task seriously. i typed "create gpu marketplace cards with hardware specs and prices per hour" and the model went and coded this ui, one shot, navy bg, glassmorphism cards, neon accent buttons, realistic pricing tiers per architecture. it even wrote a "why this looks premium" explanation under the code. for context this is a q4 quant of google's 31b dense thinking model, running on a rtx 5090 mobile 24gb in the rog scar 18 at around 15 tok/s sustained, same vram tier as a 3090 or 4090 desktop → tweet
@sudoingX · 2026-04-18T11:51
saas pricing per user makes sense until you replace the user with an agent → tweet
@sudoingX · 2026-04-18T12:07
own your compute anon. even if it means buying a single 3090, a 4090, an entry dgx spark. just one card under your desk. then prompt it freely without openai logging every thought, without anthropic indexing your work, without your queries becoming training data for the next model competing with you. compute is where thinking happens. healthy thinking needs private infrastructure. these prices are going. ram crisis is real, hbm locked through 2027, chip lead times blowing out. the 3090 you can grab today for $900 - $1,200 won't be that price next quarter. window is open right now, closing fast. → tweet
@TheAhmadOsman · 2026-04-18T00:37
Here's an old video showing Claude Code running w/ local models on my own GPUs at home > SGLang serving MiniMax-M2.1 > on 8x RTX 3090s > nvtop showing live GPU load > Claude Code generating code + docs > end-2-end on my AI cluster → tweet
@levelsio · 2026-04-18T10:53
I get 4,150,000,000 (4 billion) requests and 300,000,000 (300 million) visitors per year That'd cost me $338,913 if I wasn't on a VPS like @brian_lovin below Instead I pay $2,932/year or 115x less because I host my sites on my own VPS. Especially with AI these days it's easier than ever to host your sites yourself, and way cheaper! 100-1000x cheaper! → tweet
Developer Tools & Workflows
@badlogicgames · 2026-04-18T11:14
RT @s_streichsbier: In case people are wondering whether compaction works well in Pi. I've been using the same session for a couple of day… → tweet
@badlogicgames · 2026-04-17T21:53
RT @lucasmeijer: This is fixed (by 💕@badlogicgames) in latest Pi! use cmd-shift-up in mac ghostty to auto-scroll-up through assistant text… → tweet
@nummanali · 2026-04-18T09:26
editorzero - human / agent docs platform This is a project I've been wanting to do for a while, and it allows me do a real world test against Opus 4.7 Max It's being built fully agentically with some guidance from me Still Work in Progress*** → tweet
@badlogicgames · 2026-04-17T18:44
RT @rauchg: The hardest thing about agents and backends is durability. @workflowsdk fixes this. That LLM you're calling will go down. Th… → tweet
@nummanali · 2026-04-17T18:39
Seriously underrated tech news Email templating is hard, different for every device and app. No Open Source editor meets the bar. I love React Email, and I loved Resends Email Editor as well - they've made it fully open source! → tweet
@sudoingX · 2026-04-18T06:42
i tried gpt 5.4 for the first time through wingman and the model is shockingly good. and wingman makes it easy because you can pick from any sota model in one place, no juggling api keys, no setup hell. i fire tasks inside wingman all day. some need frontier capability and gpt 5.4 is now the answer i reach for. but my favorite part is still the self prompting autopilot mode. fire a task, walk away, wingman notifies you when it's done. → tweet
Industry Commentary & Trends
@LinusEkenstam · 2026-04-17T22:40
Today I've seen the world shift beneath my feet. Announcement after announcement, acceleration beyond comprehension. We live in a world where you sleep for 8 hours and you wake up and you've missed out. We're deep in a psychosis. Frontier labs are rushing, it's truly a race. I've seen charts of hyper scalers spending close to 1trillion in <6 years. Follow the money someone wise said. We're losing sight of staying critical in our observations. We're simply getting blinded by the shiny news infront of us. Day in and day out. There is no longer a way back. We've passed the point of no return. The orbits in AI and compute advancements are spinning faster. The acceleration can be felt. But I'm deeply concerned about our inability to stay critical. We're all in an AI psychosis stranded, amazed, confused, flabbergasted in awe. This is the time to stay awake, to take in, observe and think critically about what's happening. → tweet
@FinansowyUmysl · 2026-04-17T07:19
Raport Challenger, marzec 2026: ponad 60 000 etatów zredukowanych w USA w jednym miesiącu. AI bezpośrednią przyczyną co czwartego zwolnienia. Goldman Sachs: 300 milionów miejsc pracy globalnie narażonych na automatyzację. 25% pracy w USA i Europie może być wykonane przez AI całkowicie. CEO Microsoft AI, Mustafa Suleyman (luty 2026): "Koniec białych kołnierzyków w 18 miesięcy." CEO Anthropic, Dario Amodei: "AI zlikwiduje połowę początkowych stanowisk biurowych w ciągu 5 lat." Najbardziej zagrożeni? Prawnicy niższego szczebla. Analitycy finansowi. Księgowi. Copywriterzy. Junior programiści. Ale jest druga strona medalu: Gartner prognozuje, że 50% firm, które zwolniły ludzi "przez AI", do 2027 będzie ponownie zatrudniać na podobnych stanowiskach. To nie eliminacja. To transformacja. AI nie zabiera stanowisk. Zabiera zadania. → tweet
@jsuarez · 2026-04-18T14:54
Reinforcement Learning dev with Joseph Suarez → tweet
@jsuarez · 2026-04-18T14:51
Hopefully this is enough to finally convince a certain group of RL researchers that GPUs are not CPUs-but-faster. Looking forward to seeing this in PufferLib soon! → tweet
@kunchenguid · 2026-04-18T05:11
honest book recommendation - building things is becoming easier with AI so knowing what to build is becoming the core skill this book is seriously a must read for founders. it teaches you the basics for how to find real problems and get real feedback from others → tweet