Tech & AI: Spring 2026
💻 10 March – 5 April 2026
Period	27 days of reporting
Key model releases	GPT-5.4 mini, Gemma 4, Qwen 3.5/3.6, Nemotron Cascade 2
Top funding	OpenAI $122B round; Replit $400M at $9B
Major layoffs	Atlassian 1,600; Dell 11,000; Oracle 20,000+; Block ~40%
Security incidents	McKinsey breach (47M messages); litellm supply chain; Claude/Codex source leaks
Breakout project	Hermes Agent: 0 → 24,000 GitHub stars in 22 days

Tech & AI: Spring 2026

Summary

The period from 10 March to 5 April 2026 was defined by three converging dynamics in the technology sector: the rapid maturation of open-source AI models approaching parity with closed frontier systems, an explosion in autonomous AI agent frameworks reshaping software development, and a series of major security incidents exposing the fragility of the AI supply chain.

On the model front, Google DeepMind’s Gemma 4 family (up to 31B dense) surpassed far larger models on ELO rankings, while Alibaba’s Qwen 3.5/3.6 series achieved near-parity with Anthropic’s Opus 4.5 at a fraction of the cost. OpenAI’s GPT-5.4 processed 5 trillion tokens per day—exceeding the entire API’s yearly volume from 2025—and generated $1 billion in annualised net-new revenue, while the GPT-5.4 mini delivered a 32x efficiency gain ($11.84/task down to $0.37).

The period’s breakout project was Hermes Agent by NousResearch, which rocketed from launch to 24,000 GitHub stars in 22 days, becoming the third fastest-growing repo on GitHub and the fifth most-used AI agent globally. The broader AI agent ecosystem—including Cursor, Claude Code, OpenAI Codex, and the new Hermes Agent—saw rapid competitive leapfrogging with no defensible positions emerging.

Security dominated headlines: McKinsey’s internal AI chatbot “Lilli” was breached, exposing 47 million chat messages and 728,000 client records via trivial SQL injection. A supply chain attack on the litellm PyPI package (97 million monthly downloads) exfiltrated SSH keys, cloud credentials, and crypto wallets—discovered only because the attacker’s memory leak bug caused a system crash. Both Claude Code and OpenAI Codex had their source code leaked within the same 24-hour window.

Financially, OpenAI closed a landmark $122 billion funding round, while the industry shed over 30,000 jobs across Atlassian, Dell, Oracle, Block, and rumoured Meta cuts—all explicitly linked to AI-driven restructuring.

AI model releases & benchmarks

Frontier and near-frontier models

Model	Company	Date	Key specs
GPT-5.4	OpenAI	~16 Mar	5T tokens/day; $1B annualised net-new ARR
GPT-5.4 mini	OpenAI	17 Mar	32x efficiency over GPT-5.2; $0.37/task (from $11.84)
Gemma 4 (31B Dense)	Google DeepMind	2 Apr	Surpasses Qwen 122B and DeepSeek v3.2 on ELO
Gemma 4 (26B MoE, 4B active)	Google DeepMind	2 Apr	Parity with Kimi K2.5 (1.1T params) at 35x smaller; runs on mobile
Qwen 3.5 27B Dense	Alibaba	26 Mar	256K context; fits on RTX 5090; “Claude Sonnet 4.6 quality at home”
Qwen 3.6-Plus	Alibaba	31 Mar	Near-parity with Opus 4.5; wins Terminal-Bench 2.0; tied on SWE-bench
GLM-5.1	Zhipu	21 Mar	Open-source; long-horizon agentic engineering (week-scale tasks)
MiniMax M2.7	MiniMax	31 Mar	Open-sourced; “closest to fully local Claude Code + Opus 4.6”
Nemotron 3 Super 120B	NVIDIA	11 Mar	MoE (12B active); 1M context; open weights/datasets
Nemotron Cascade 2	NVIDIA	25 Mar	Mamba-2 arch; 187 tok/s on RTX 3090; flat perf 4K–625K context
Trinity-Large-Thinking	—	2 Apr	#2 PinchBench; Apache 2.0; ~96% cheaper than Opus

Specialised models

• Cohere Transcribe (27 Mar): 2B-parameter multilingual (14 languages) speech recognition; Apache 2.0; handles poor audio quality.
• Voxtral-4B-TTS (Mistral, 26 Mar): Open-source text-to-speech on Hugging Face.
• NVIDIA Kimodo (23 Mar): Text-to-3D motion synthesis; 700 hours of mocap data; free on Hugging Face.
• Google Flood Forecasting Model (12 Mar): Predicts urban flash floods up to 24 hours in advance.
• PrismAudio V2A (Alibaba, 24 Mar): RL-powered video-to-audio model.
• Mr. Chatterbox (30 Mar): Victorian-era language model trained on 28,000+ British texts (1837–1899).

Performance degradation

All major frontier models—Opus 4.6, ChatGPT 5.4, and Codex 5.3—experienced significant performance degradation during this period. Flutter development was particularly affected, with non-compiling tests and false passing claims. Claude Opus 4.6 was reported entering circular reasoning loops, while ChatGPT 5.4 created overly complex solutions for simple problems.

AI agents & coding tools

Hermes Agent (NousResearch)

The breakout project of the period. Launched to 24,000+ GitHub stars in 22 days, making it the #3 fastest-growing GitHub repo and the 5th most-used AI agent globally:

• 263 PRs merged in 6 days across 6 major releases (v0.4.0 through v0.7.0)
• 11 per-model tool call parsers (Qwen, DeepSeek, LLaMA, Mistral, GLM)
• Local-first, privacy-preserving, zero per-token costs
• v0.7.0 features: pluggable memory (7 providers from local SQLite to knowledge graphs), credential pools with automatic API key rotation, Camofox anti-detection browser, secret exfiltration blocking
• “GODMODE” skill for automatic model jailbreaking (27 Mar)
• Background self-improvement capability; OpenAI API-compatible SDK
• Community: 4,000+ members; ranked 6th largest AI app on OpenRouter

Coding assistant landscape

The AI coding tool market saw rapid competitive leapfrogging with no defensible positions:

• Cursor AI: Released first custom model (20 Mar); launched Cursor 3 with Composer 2 (3 Apr); discovered to use Moonshot AI’s Kimi foundation model (Chinese lab powering Western tools); trained Composer to self-summarise via RL, reducing compaction errors by 50%.
• Claude Code: Added scheduled automation, UI mockup generation via AskUserQuestion; mobile SSH productivity workflow gained popularity; source code leaked via sourcemaps (31 Mar); Anthropic increasingly restricting third-party harness access, driving migration to alternatives.
• OpenAI Codex: Sam Altman reported “hardcore builders” switching with “very fast” usage growth; subagent support for parallel workflows (17 Mar); moved to usage-based API pricing (3 Apr); source code also leaked (31 Mar).
• OpenClaw: Telegram integration; Microsoft Foundry MCP browser-driving support; but users migrating to Hermes Agent citing slow gateways and broken tool calls.

Autonomous development paradigm

• Karpathy’s Autoresearch (10 Mar): Autonomous AI agent improved nanochat training efficiency by 11% over 2 days; ~20 additive improvements across 700 autonomous experiments. Karpathy predicted this approach will become standard at frontier labs.
• Meta Hyperagents (26 Mar): 200 autonomous agents handling production code shipping.
• Karpathy’s vision (26 Mar): Fully autonomous DevOps agents—“tell agent ‘build menugen’ → deployed web app with zero manual steps.”
• levelsio demo (28 Mar): 24-minute MVP build (personalised bedtime story app) using Claude Code, Grok 4.1 TTS, Gemini, and Stripe integration.

Open-source momentum

Open-source AI models approached parity with closed frontier systems during this period:

• Gemma 4 31B surpassed Qwen 122B and DeepSeek v3.2 on ELO rankings
• Qwen 3.6-Plus achieved near-parity with Opus 4.5 on SWE-bench and Terminal-Bench
• Trinity-Large-Thinking reached #2 on PinchBench at 96% less cost than Opus, with Apache 2.0 weights
• MiniMax M2.7 open-sourced, described as closest to fully local Claude Code experience
• llama.cpp reached 100,000 GitHub stars (30 Mar)
• American open-source labs predicted to “narrow the gap with closed models in 2026–2027”

Key tension: Anthropic criticised for restrictive licensing on skills (vs. Apache 2.0 for most Codex skills) and for blocking third-party access through OpenClaw. Users migrating to Codex, Droid, Kimi CLI, and OpenCode alternatives.

Hardware & infrastructure

NVIDIA GTC highlights

• DLSS 5: Described as the “ChatGPT moment for computer graphics”—breakthrough photorealism. GPU-based real-time AI upscaling where every pixel is AI-generated at runtime.
• OpenShell: Free, open-source (Apache 2.0) sandboxed runtime for AI agents.
• Jensen Huang framed AI factory throughput as the key benchmark for infrastructure investment; engineers earning $500k should consume $250k+ in AI tokens.
• NVIDIA: up 1,000x since 2009 ($4B to $4T+ valuation) versus Intel’s 2x gain.

Consumer hardware

• Apple Neo MacBook (15 Mar): $599, modular design, swappable ports/keyboards, standard screws—“most repairable MacBook ever.”
• Apple M5 Pro/Max (23 Mar): Three Thunderbolt 5 ports enabling RDMA clustering of up to 4 MacBooks with single-digit microsecond latency for tensor parallelism. Supports 1-trillion parameter models at 70+ tok/s with near-linear scaling.
• RTX 3090: Emerged as the “undisputed” workhorse GPU for local inference. A single $900 RTX 3090 running Qwen 3.5 27B dense repeatedly outperformed 120B+ MoE models on $70K enterprise hardware in real-world coding tasks.

Infrastructure projects

• Stargate Michigan Data Center (27 Mar): First steel beams erected; OpenAI with Oracle and Related Digital.
• Tinygrad Modular Datacenter (10 Mar): Truck-deployable shipping containers; 600kW at <$0.05/kWh; first unit $3M; target 20 exaflops.
• Tenstorrent Cluster (27 Mar): >1TB VRAM, 3TB DDR5, 32TB SSD.
• Google GKE (17 Mar): Replaced etcd with Cloud Spanner for up to 65,000-node clusters.
• Ollama: Upgraded to NVIDIA B300 hardware; launched $200/year Pro plan; MLX integration for 2.2x speed on Apple silicon.
• Code.Storage (25 Mar): New Git provider for AI/machine-generated repos (GitHub strained by ~230 AI repos/day).

Quantisation breakthroughs

• Google TurboQuant (25 Mar): KV cache compression reducing memory by minimum 6x; already in MLX framework.
• Qwen 3.5 35B: 20% size reduction with ~1% performance loss; fits in 4-bit on 24GB VRAM.
• 12GB VRAM identified as viable minimum threshold for serious local inference.

Cybersecurity incidents

McKinsey “Lilli” chatbot breach (13 Mar)

Internal chatbot trained on 100,000 documents, used by 70% of 45,000 employees (500,000 monthly prompts). Exposed: 47 million chat messages (strategy/M&A data), 728,000 confidential client records, 57,000 user accounts, 95 system prompts. Root cause: publicly exposed API documentation with 22 unauthenticated endpoints vulnerable to SQL injection. Patched after discovery.

litellm PyPI supply chain attack (24 Mar)

Versions ≥1.64.0 exfiltrated SSH keys, AWS/GCP/Azure credentials, Kubernetes configs, git credentials, API keys, shell history, crypto wallets, SSL private keys, CI/CD secrets, and database passwords. The package has 97 million monthly downloads, and transitive dependencies (notably DSPy) were also compromised. The attack was discovered only because the attacker’s memory leak bug caused a system crash. Karpathy called supply chain attacks “the scariest thing imaginable in modern software.”

Source code leaks (31 Mar)

Both Anthropic’s Claude Code and OpenAI’s Codex had their source code exposed within the same 24-hour window. Claude Code was leaked via sourcemaps (~512,000 lines); thousands of forks appeared within hours. Anthropic sent DMCA takedown notices—ironic given it had issued similar takedowns against open-source forks 336 days earlier.

Other incidents

• Cline VS Code extension (19 Mar): Compromised via prompt injection installing “OpenClaw” malware; ~4,000 machines affected.
• npm axios (31 Mar): Supply chain attack on package with 300 million weekly downloads; suspicious imports from googleworkspace/cli discovered.
• Chrome extension malware (1 Apr): Criminals acquiring popular extensions, injecting code, compromising cookies and localStorage.
• SSH brute force (18 Mar): Server documented 165,225+ attempts over 90 days.

Corporate moves & funding

Major funding

Company	Amount	Notes
OpenAI	$122 billion	Largest AI funding round ever (1 Apr). Offering PE firms 17.5% guaranteed minimum return + early model access.
Replit	$400M at $9B	Launched Replit Agent 4 (11 Mar)
Yann LeCun’s lab	$1.03 billion	New AI research lab with global mission (10 Mar)
OpenAI Foundation	$1B first-year spend	Focus: biological threats, economic disruption, societal effects (24 Mar)

Acquisitions

• Netflix acquired InterPositive (Ben Affleck’s AI film startup) for up to $600M (12 Mar). Netflix’s largest tech acquisition; AI-augmented post-production.
• OpenAI acquired Astral (creators of uv Python package manager and ruff linter, 19 Mar). Team integrating into Codex division.
• OpenAI acquired TBPN (media/communications platform, 2 Apr). Signals expansion into AI-native content production.

Service shutdowns

• OpenAI Sora video generation platform and API discontinued (25 Mar).
• Meta Horizons Worlds metaverse social platform shut down (19 Mar).

Layoffs & workforce shifts

Over 30,000 jobs were shed in a four-week period, all explicitly linked to AI-driven restructuring:

Company	Cuts	Date	Notes
Oracle	20,000+	1 Apr	—
Meta	~15,000 (rumoured)	14 Mar	20% reduction to fund AI investments
Dell	11,000	19 Mar	—
Atlassian	1,600 (10%)	12 Mar	Cited AI impact on business
Block	~40% headcount	20 Mar	Framed as AI automation

PwC benchmarked OpenAI’s finance team at 20% of typical company size. Jensen Huang stated engineers earning $500k should consume $250k+ in AI tokens. Developers in the Polish IT industry reported AI tools increasing workload due to layoff fears rather than improving quality of life.

The code review crisis was highlighted by Gene Kim and Jez Humble: AI-generated code volume is overwhelming traditional review processes, raising questions about quality assurance, accountability, and career development.

Developer tools & platforms

• Ollama v0.18.1 (18 Mar): Web search/fetch for OpenClaw; non-interactive headless mode for CI/CD. Upgraded to NVIDIA B300 cloud infrastructure. Launched $200/year Pro plan. MLX integration for 2.2x Apple Silicon speed.
• Ghostty Terminal: Progressing through libghostty C API; builds/tests on Windows CI (24 Mar); surpassed Terraform’s GitHub star count (23 Mar).
• Google Stitch (19 Mar): AI-native UI/design tool positioned as Figma competitor; integrated with Gemini.
• EXO 1.0.69 (28 Mar): Qwen 3.5 support, continuous batching, M5 Pro/Max compatibility. Distributed inference across Mac Studios.
• ArrowJS 1.0 (23 Mar): Open-source UI framework for coding agents, no compiler/build step.
• OpenScreen (2 Apr): Open-sourced screen capture; 8,400+ GitHub stars; MIT licensed, cross-platform.
• Transformers.js v4 (31 Mar): WebGPU backend for browser and Node.js ML inference.
• Python 3.15 JIT compiler (24 Mar): Back on schedule.
• .NET Aspire (19 Mar): ~930 PRs merged; TypeScript support added.
• Chrome 146 (14 Mar): Enables local AI model exposure with single toggle.

Drone & defence tech

Ukraine’s drone technology became a major export and diplomatic asset (see also Russo-Ukrainian War: Spring 2026):

• 10,000 Merops AI-enabled interceptor drones (developed in Ukraine) deployed by the U.S. military against Iranian Shaheds; ~$15K each ($5K at scale).
• 201–228 Ukrainian counter-drone specialists deployed to Gulf states.
• Saudi Aramco in talks with SkyFall, Wild Hornets, and Phantom Defense.
• Japan exploring purchase of Ukrainian combat drones with technology transfers.
• Niantic Spatial (Pokemon Go parent, 16 Mar): Powers 1,000 autonomous Coco Robotics delivery vehicles with centimetre-accurate GPS-free navigation from 30+ billion user-contributed images.

AI in science & medicine

• Personalised mRNA cancer vaccine (14–15 Mar): Australian developer Paul Conyngham designed a personalised mRNA cancer vaccine for his rescue dog using ChatGPT and AlphaFold. Budget: $3,000. No biology background. Result: tumour size reduction. Sam Altman highlighted as potential company opportunity.
• Medical diagnosis (21 Mar): Patient used ChatGPT to identify cancer treatment options after physicians exhausted standard protocols.
• ICML 2026 (18 Mar): 497 submissions desk-rejected for violating AI reviewer use policy—largest known academic enforcement action. A previous PufferLib paper had been rejected based on hallucinated typos from an AI reviewer.
• AI research integrity (13 Mar): Papers accused of falsified results—Pong scores exceeding the maximum possible value of 21.
• John Deere (30 Mar): 92% daily AI engagement, 300 parallel AI experiments in agricultural tech.
• Cargill CarVe (22 Mar): Machine vision recovering 55 million lbs of meat annually (~$200M value).

Local inference revolution

A recurring theme was the surprising competitiveness of local inference on consumer hardware versus expensive enterprise infrastructure:

The dense vs. MoE surprise

Qwen 3.5 27B dense running on a single $900 RTX 3090 repeatedly outperformed 120B+ MoE models on $70K enterprise nodes (2x H200 NVL) in real-world autonomous coding tasks. Dense models (every parameter active per token) showed unexpected strength in agent-based coding where architectural coherence matters more than raw parameter count.

Recommended inference engines (by use case)

• llama.cpp: CPU/GPU/Mac portability (100K GitHub stars; recommended over Ollama for agent workloads due to abstraction overhead)
• MLX: Apple Silicon unified memory (recommended over llama.cpp for M-series)
• ExLlamaV2: Single RTX optimisation
• vLLM: Production serving default
• SGLang: Complex multi-model infrastructure
• TensorRT-LLM: Maximum NVIDIA performance

Key benchmarks

• Qwen 3.5 27B Dense (Q4_K_M): 35 tok/s on RTX 3090, 262K context with zero degradation
• Qwen 3.5 35B MoE: 112 tok/s on RTX 3090, only 3B params active
• NVIDIA Nemotron Cascade 2: 187 tok/s on RTX 3090 (IQ4_XS), 625K context
• Gemma 4 31B on MacBook Pro M4: TTFT 5.68s, prompt 652 tok/s, decode 40 tok/s
• M5 Pro/Max RDMA cluster (4 MacBooks): 1T-parameter models at 70+ tok/s

VRAM landscape

Community survey: 8–12GB (31.6%), 24GB (34.5%), 48GB+ (21.5%). Memory bandwidth (not VRAM size) identified as the true local LLM bottleneck.

Policy & regulation

• Anthropic CEO (21 Mar): “Disagreeing with the government is the most American thing in the world.”
• NVIDIA CUDA moat: Millions of developers over 20 years; installed on 1B+ devices (27 Mar).
• Super Micro (20 Mar): Implicated in $2.5B NVIDIA GPU smuggling scheme.
• Apple Siri (25 Mar): Reportedly switching to Google Gemini foundation model, signalling Apple discontinuing internal AI model development.
• McKinsey at ASML (21 Mar): Consulting involvement raised supply chain influence concerns in semiconductor manufacturing.
• DoorDash Tasks (20 Mar): App using Dashers to film chores for AI robotics training data.

Emerging trends

AI adoption metrics

• AI adoption entering “early majority” phase (34% of users); predicted to double within 12 months (13 Mar).
• 78% of Gen-Z can identify AI-generated images; 40% click-through-rate drop observed using AI imagery (18 Mar).
• Apple App Store flooding with low-quality AI-generated apps (24 Mar).
• Token cost to build a production feature is now lower than the meeting cost to discuss building it (13 Mar).

Developer sentiment

• “I don’t enjoy coding anymore because it’s so easy with AI” (12 Mar).
• The basic unit of development shifting from file to agent (11 Mar).
• MCP protocol critique: “AI doesn’t need abstractions because they can use APIs directly” (12 Mar).
• Dr. Erik Meijer’s Universalis programming language (12 Mar): forces LLMs to generate mathematical proofs of safety before execution. “We are going to be the last generation of developers to write code by hand.”
• Solo developer “Ben” reached $4.4M ARR in under 2 months, described as on track for first billion-dollar solo founder (18 Mar).

Platform dynamics

• Vendor lock-in pattern: API restrictions framed around safety interpreted as platform capture (5 Apr).
• Open-source counter-movement: “bring your own device/data” philosophy with exponential OpenRouter usage growth.
• “Cognitive security” concern driving local LLM adoption.
• GitHub approaching “zero nines availability” from AI agent activity strain (~230 machine-generated repos/day).
• OpenAI pursuing vertical integration (chips, healthcare, infrastructure); Anthropic maintaining focused model/API approach.

Upcoming models (2026 watch list)

MiniMax M3 (multimodal), NVIDIA Nemotron 3 Ultra (~500B parameters), Kimi K3, DeepSeek V4, GLM-5.1 full release, GPT-Image-2 (leaked on arena leaderboards under codenames maskingtape-alpha, gaffertape-alpha, packingtape-alpha).

Condensed timeline

Date	Key events
10 Mar	Karpathy’s Autoresearch: 11% efficiency gain via 700 autonomous experiments. Yann LeCun raises $1.03B for new lab. Tinygrad modular datacenter ($3M, 20 exaflops target).
11 Mar	NVIDIA Nemotron 3 Super 120B (1M context). Replit: $400M at $9B. Claude Code adds scheduled automation. TADA open-source TTS.
12 Mar	Netflix acquires InterPositive ($600M). Atlassian lays off 1,600. Ukraine opens battlefield data for AI training. Universalis: “last generation to write code by hand.”
13 Mar	McKinsey “Lilli” breach: 47M messages, 728K client records via SQL injection. U.S. deploys 10,000 Ukrainian Merops drones. AI adoption at 34% “early majority.”
14 Mar	Ollama upgrades to NVIDIA B300. Personalised mRNA cancer vaccine ($3K, ChatGPT + AlphaFold). Chrome 146 local AI toggle. Meta rumoured 20% cuts.
15 Mar	Apple Neo MacBook ($599, modular). Opus 4.6 / GPT-5.4 / Codex 5.3 all degrading. Open-source radar at 95% cost reduction.
16 Mar	GPT-5.4: 5T tokens/day, $1B net-new ARR. Niantic Spatial: 1,000 autonomous delivery robots. NVIDIA up 1,000x since 2009. Code review crisis emerges.
17 Mar	GPT-5.4 mini: 32x efficiency ($0.37/task). NVIDIA DLSS 5 “ChatGPT moment for graphics.” GKE replaces etcd with Spanner. Codex subagents for parallel workflows.
18 Mar	ICML 2026: 497 papers desk-rejected for AI reviewer use. Nemotron 3 Nano 4B for edge. Shahed drone flies 5 km from Polish border. Solo dev “Ben” at $4.4M ARR.
19 Mar	OpenAI acquires Astral (uv + ruff). Google launches Stitch (Figma competitor). Cline VS Code compromised (~4K machines). Dell lays off 11,000. Meta shuts Horizons Worlds.
20 Mar	Cursor releases first custom model (routes through Kimi/Moonshot). Block cuts ~40%. Super Micro: $2.5B GPU smuggling. Telegram added to Claude. DoorDash films chores for robotics data.
21 Mar	GLM-5.1 open-sourced. Halter (AI cow collars) valued at $2B+. Patient uses ChatGPT for cancer treatment. Anthropic CEO on government dissent.
22 Mar	MiniMax M2.7 open weights announced. Hermes Agent: 10K stars, 8+ day continuous runs. GPU supply chain shortages. Cargill CarVe: $200M meat recovery. Terrafab rumour (Tesla/xAI/SpaceX chips).
23 Mar	Apple M5 Pro/Max: RDMA clustering, 1T-param models at 70+ tok/s. NVIDIA Kimodo: text-to-3D motion. Cursor revealed using Chinese Kimi model. Framework price reductions.
24 Mar	litellm supply chain attack: SSH keys, cloud creds exfiltrated (97M monthly downloads). Hermes Agent v0.4.0: 300 PRs in one week. Nemotron Cascade: 3B active, IMO gold-medal math. OpenAI Foundation ($1B). Lace Lithography: 10x Moore’s Law extension claim.
25 Mar	Nemotron Cascade 2: 187 tok/s on RTX 3090. Apple Siri switching to Gemini. OpenAI Sora discontinued. TurboQuant: 6x memory reduction. Code.Storage for AI-generated repos.
26 Mar	Qwen 3.5 27B: “release of 2026 so far.” Hermes Agent: 13.3K stars. Karpathy proposes fully autonomous DevOps agents. OpenAI erotic features shut down in 48h.
27 Mar	Stargate Michigan: first steel beams. Cohere Transcribe (2B, 14 languages). Hermes Agent GODMODE for model jailbreaking. mRNA vaccine highlighted by Altman. Apple rumoured reconsidering SwiftUI.
28 Mar	Gemini 3.1 Flash Live API. Qwen 3.5 27B: 35 tok/s on RTX 3090, 262K context. levelsio: 24-min MVP build. RLHF critique: optimises for confidence not accuracy.
29 Mar	Hermes Agent v0.5.0: browser automation. Qwen 9B builds 2,699-line game on first try. Nemotron Cascade 2 fails coding coherence tests. llama.cpp > Ollama for agents.
30 Mar	llama.cpp reaches 100K GitHub stars. John Deere: 92% daily AI engagement. Hermes Agent: 3,000+ members. Claude Code integrates with Codex.
31 Mar	Claude Code + Codex source code leaked in same 24h window (~512K lines). npm axios supply chain attack (300M downloads). MiniMax M2.7 open-sourced. Transformers.js v4 (WebGPU). Qwen 3.6-Plus preview.
1 Apr	OpenAI closes $122B funding round. Oracle lays off 20,000+. Hermes Agent ranked 6th AI app on OpenRouter. Tinygrad Mac eGPU support. Chrome extension malware wave.
2 Apr	Google DeepMind Gemma 4 (31B surpasses Qwen 122B on ELO). Trinity-Large-Thinking: #2 PinchBench, 96% cheaper than Opus. OpenAI acquires TBPN. OpenScreen open-sourced (8.4K stars).
3 Apr	Gemma 4 26B MoE matches 1.1T-param Kimi K2.5 at 35x smaller. Cursor 3 launches. Hermes Agent: #3 fastest-growing GitHub repo. Vibe Jam 2026 ($35K prizes). mesh-llm open-sourced by Block.
4 Apr	Hermes Agent v0.7.0: 24K stars (from 13K in 2 days). GPT-Image-2 leaked on arena. Ollama Cloud ($20/mo). RTX 3090 ($900) outperforms $70K H200 nodes in coding. Vibe Jam: participants as young as 10.
5 Apr	Anthropic blocks third-party access. Gemma 4 E2B runs on iPhone 17 Pro. Block open-sources Goose agent (150+ MCP servers). PwC: OpenAI finance team at 20% of normal size.

Sources

Compiled from daily tech/AI/IT intelligence reports published at waclaw.online/raport2/tech/, covering 10 March – 5 April 2026 (27 daily reports). Individual reports available for each date.