Executive Summary
The past 24 hours were dominated by Google I/O announcements and their aftermath, with Gemini 3.5 Flash receiving mixed reception—praised for cost-efficiency but criticized for disappointing output quality. OpenAI made a bold strategic move by offering $2M in API credits to every Y Combinator startup, signaling an aggressive land-grab amid expectations of capacity constraints. The open-source agent ecosystem rapidly matured, with Hermes Agent (NousResearch/Teknium) shipping multiple upgrades including xAI integration, BrowserBase skills, and an F-Droid native app. On the security front, a supply-chain compromise of the mistralai npm package prompted hardening releases, and reports emerged of 4,000 GitHub private repositories being stolen and listed for sale.
Key Events
- OpenAI offers $2M in tokens to every YC startup, signaling aggressive developer ecosystem capture and anticipating a capacity-constrained future. → link
- Gemini 3.5 Flash launches to mixed reviews — praised as cost-competitive with GPT 5.5 but criticized for quality gaps versus Gemini 3.1 Pro and disappointing benchmark performance. → link
- Gemini Omni debuts multimodal video generation capabilities, including an alphabet-matching Turing test demo. → link
- Hermes Agent ships 4 major speed upgrades plus xAI web search integration, BrowserBase domain skill hub, and a native F-Droid app. → link
- Supply-chain attack on
mistralainpm package triggers security hardening across the ecosystem; a malicious payload was also found inaidenybai/million. → link - 4,000 GitHub private repositories reportedly stolen and listed for sale at $50K by a hacker group. → link
- NVIDIA open-sources SANA-WM, a camera-conditioned world model (2.6B params, Apache 2.0) generating 60s of 720p video on a single RTX 5090. → link
- Cursor Composer 2.5 launch — nearly matching opus 4.7 max quality at ~10x lower cost, building pressure on Anthropic pricing. → link
- Tinygrad merges first assembly backend (DEV=CPU:X86), advancing toward kernels that outperform LLVM/PTX-layer approaches. → link
- Qwen3.6-27B-PRISM-PRO-DQ released on HuggingFace, trained for long-horizon multi-turn agentic work. → link
Analysis
Patterns: The competitive landscape is bifurcating between closed-model providers racing to lock in developers (OpenAI's YC credit blitz, capacity pre-commits) and open-source agent tooling rapidly maturing (Hermes, Qwen, SANA-WM, Tinygrad). Token economics dominated discourse — from OpenAI's capacity-constrained warnings to complaints about Claude Code pricing/limits to Cursor's disruptive cost positioning.
Escalation: Supply-chain attacks on npm packages are escalating. The mistralai compromise and the million package malicious payload, combined with the GitHub repo heist, underscore that the intersection of AI tooling and package management is a growing attack surface. Multiple providers are responding with hardening measures.
Google's positioning problem: Google I/O generated buzz but also frustration — Gemini 3.5 Flash was called "disappointing," the new Antigravity IDE was criticized for using Electron over Flutter, and demos reportedly used iPhones. This recurring pattern of strong announcements followed by execution gaps continues to dog Google.
What to watch next: Gemini 3.5 Pro benchmarks (expected to be significantly stronger); whether OpenAI's $2M YC credit offer triggers competitive responses from Anthropic and Google; the trajectory of supply-chain attacks targeting AI-adjacent npm packages; and whether community-built agent frameworks like Hermes can sustain their rapid integration pace.
Tweet Feed
AI Model Releases & Benchmarks
@TheAhmadOsman · 2026-05-20T12:59
Gemini 3.5 Flash is disappointing. Why does Google keep fumbling their models? Genuinely baffled by this → tweet link
@MengTo · 2026-05-20T11:02
I've created 50 designs with Gemini 3.5 Flash. Here are my thoughts: - As good as GPT 5.5 but way cheaper. - Worse than Gemini 3.1 Pro in term of design complexity like layout grids, responsiveness, scroll behaviors, webgl, animations, etc. - Omni looks good for video edits, but lacks emotion compared to Seedance 2.0. - Omni could be a game changer for web design if good with typography. Can't wait for Gemini 3.5 Pro. → tweet link
@FinansowyUmysl · 2026-05-19T19:05
Wchodzimy w nowy cykl publikacji nowych modeli. Gemini 3.5 flash. Bardzo dobre benchmarki. Gemini 3.5 pro może być bardzo dobry. → tweet link
@TrungTPhan · 2026-05-19T23:13
Google Omni did the "create the alphabet and match each letter with an image starting with that letter" Turing Test…for video → tweet link
@FinansowyUmysl · 2026-05-20T05:01
Cursor (edytor z wbudowanym AI), niedawno zaczął trenować własne modele. Ich najnowszy model Composer 2.5 jest prawie tak dobry jak opus 4.7 max czy openAI xhigh, ale... jest nawet 10x tańszy! → tweet link
@Ex0byt · 2026-05-19T21:31
And... Ladies and Gentlemen: Qwen3.6-27B-PRISM-PRO-DQ - enjoy! → tweet link
@Ex0byt · 2026-05-19T21:21
The drafter is shipped and on HuggingFace, as promised. I trained it for extremely long-horizon, deep multi-turn agentic work, more training only makes it even better. → tweet link
@victormustar · 2026-05-20T14:09
[NEW] Hugging Face Dataset Leaderboard: you can now filter benchmark results by model parameter range! For example: find the best model under 32B on SWE-bench. → tweet link
OpenAI Strategy & Token Economics
@sama · 2026-05-20T03:00
i am excited to see what will happen with tokenmaxxing startups, both for how they work internally and the products they can build. openai offered to invest $2M in tokens into every startup in the current yc batch. happy building! → tweet link
@gdb · 2026-05-19T22:25
we are offering discounted tokens and certainty on capacity availability in exchange for 1-3 year commits. we expect that the world will feel increasingly capacity constrained for the next while, as models continue to get much more useful. → tweet link
@sama · 2026-05-19T19:59
customers are increasingly asking us for certainty on capacity. as models get better, we expect that the world will be capacity-constrained for some time. we are offering discounted tokens for 1-3 year commits. → tweet link
@gdb · 2026-05-20T04:01
openai offering to invest $2M in API credits in every @ycombinator startup in the current batch. compute for powering the next generation of startups. → tweet link
@jezell · 2026-05-20T06:38
If you thought Salesforce licenses were expensive, you should try paying for tokens. → tweet link
@LinusEkenstam · 2026-05-20T00:19
Tokenmaxxing 🫡 → tweet link
Agent Ecosystem & Developer Tools
@Teknium · 2026-05-20T18:23
Access skills made directly for the websites your agents use with BrowserBases' domain skillshub integration in Hermes Agent! Access early with
hermes update→ tweet link
@Teknium · 2026-05-20T03:34
4 big upgrades to Hermes Agents speed today.
hermes updateto get moving faster now → tweet link
@Teknium · 2026-05-19T23:19
The canonical guide to using X with Hermes 👀👀 → tweet link
@Teknium · 2026-05-20T07:35
RT @Jaaneek: You can now do xAI web_search in @NousResearch Hermes Agent → tweet link
@gdb · 2026-05-19T23:40
codex app is very good → tweet link
@sudoingX · 2026-05-20T04:43
cursor cli is so fast it's almost disrespectful. this is my frontier driver now. → tweet link
@sudoingX · 2026-05-20T04:49
just woke up to my aws credits at zero. nothing says good morning like a billing alert. the cloud giveth and the cloud taketh away. this is why i own my compute. back to the spark. → tweet link
@sudoingX · 2026-05-20T10:03
the part of the foundation people miss is what changes after you set it up. it does not make you faster, it makes you less busy. [...] the foundation buys you calm, and calm is what compounds. → tweet link
@steipete · 2026-05-20T10:08
Can't recommend @cotypist enough. Autocomplete everywhere. → tweet link
@jsuarez · 2026-05-20T17:09
Reinforcement learning research with Joseph Suarez → tweet link
Security & Supply Chain
@badlogicgames · 2026-05-20T14:37
People of libgdx. Supply-chain hardening release. Last week the mistralai package got shai huluded, which gave us a little scare (we were not affected, due to pinning). Starting today, we have the following safe-guards in place: - cut down dependencies to the absolute minimum - direct external deps are pinned - the CLI ships an npm shrinkwrap for transitive deps - pi update --self disables lifecycle scripts - new dependency lifecycle scripts require explicit review - lockfile changes are blocked pre-commit unless explicitly allowed - scheduled npm audit + registry signature checks run on GitHub - 2fa releases, obviously → tweet link
@badlogicgames · 2026-05-20T18:18
RT @aidenybai: Security update: We found a malicious payload in aidenybai/million and reverted it. npm or published users are NOT affected… → tweet link
@RydMike · 2026-05-20T18:31
RT @DODOREACH: A HACKER GROUP JUST STOLE 4,000 OF GITHUB'S OWN PRIVATE REPOSITORIES.. PUT THEM UP FOR SALE FOR $50,000... → tweet link
@gdb · 2026-05-19T23:10
the team is doing great work on agentic security → tweet link
@gdb · 2026-05-19T19:32
SynthID for checking if an image was generated by OpenAI: → tweet link
Claude Code & Developer Experience Issues
@sudoingX · 2026-05-20T11:58
claude code usage is cooked. i'm using less than last week and somehow my weekly limit is getting nuked faster. already at 80% with a lighter workload, while claude code is slower and buggier than ever. → tweet link
@sudoingX · 2026-05-20T10:41
claude code is unusably slow and buggy. plus they said they doubled the limits and somehow i'm running out on even less workload than normal. → tweet link
Google I/O & Antigravity IDE
@jezell · 2026-05-19T19:53
New Antigravity isn't an IDE anymore, but Google still chooses Electron over Flutter for it. → tweet link
@ASalvadorini · 2026-05-19T20:54
Coded this wanna be random #Flutter web app with Antigravity 2.0 + Gemini 3.5 Flash (high), the experience was horrible for me, I blame both, now I will happily go back to Antigravity IDE 😇 → tweet link
@ASalvadorini · 2026-05-19T19:22
Antigravity (left) vs Antigravity IDE (right). Google product idea at its essence. → tweet link
@jezell · 2026-05-20T03:26
This is the biggest news out of Google I/O → tweet link
@LinusEkenstam · 2026-05-20T17:10
Having a good laugh right now. Google I/O is delivering on so many fronts. → tweet link
Open Source & Hardware
@victormustar · 2026-05-20T07:49
it's open source time, with a real leap for world models 🎉 NVIDIA's SANA-WM: a camera-conditioned world model that fits on one GPU. 60s of 720p in 34s on a single 5090 - 2.6B params and Apache 2.0! → tweet link
@tinygrad · 2026-05-19T19:55
Our first assembly backend, DEV=CPU:X86 is merged! Instruction selection + register allocation added, and it's all visible with VIZ=1. This is the path to kernels that outperform everyone playing at the LLVM/PTX layer. → tweet link
@Teknium · 2026-05-20T16:18
RT @JeffSte17427059: @Teknium @NousResearch I'm proud to announce Hermes app native on F-Droid → tweet link
AI Commentary & Industry Analysis
@TrungTPhan · 2026-05-20T16:28
RT @bearlyai: Demis on why world models are his longest standing passion and explains benefits vs. language models: "I think language m…" → tweet link
@juliarturc · 2026-05-20T16:06
Task agnostic robots truly give me Black Mirror vibes. Like V-JEPA 2-AC from Meta, which was trained on unlabeled videos, without any task-specific training or reward... you just show the little dude a picture of what you want and it'll do it for you. → tweet link
@juliarturc · 2026-05-20T05:41
Why harvest data only from your own employees like the Zuck when you can harvest it from 200 companies → tweet link
@TheAhmadOsman · 2026-05-20T12:59
Good one by SemiAnalysis, Cerebras is known for these things btw, I don't trust them, their numbers, or their hardware / offerings. There's a person in the Local AI space that they pay to push their narrative btw → tweet link
@LinusEkenstam · 2026-05-20T14:42
This is the narrative now: Train on your employees, tell them in plain sight, rinse and repeat. I feel this fear mongering is not really what we need now, layoffs will always happen, but the narrative here makes it all look quite bad. → tweet link
@TrungTPhan · 2026-05-19T21:40
Complete 180-degree turn on AI at commencement speeches. This year, student booing every mention of it. Last year, students showing off how ChatGPT got them the degree. → tweet link
@levelsio · 2026-05-19T19:01
I'm just as surprised nobody in AI voice tech has realized a voice needs background and environmental noise to sound realistic. Even @ElevenLabs the leader in voice AI can not produce voice with background noise, or environment reverb sound. AI voices are always going to sound non-passable as human if they don't have that. → tweet link
@mraleph · 2026-05-20T07:24
LLMs whenever I try to use them for anything non-trivial → tweet link