Software Engineer. Interested in programming languages, tooling, LLMs.
Maceió, Brasil
Joined January 2010
- Tweets4K
- Following534
- Followers289
- Likes14.3K
Do I know someone hiring an iOS engineer? A friend of mine is looking for a job, remote, US timezone. If so, ping me.
Johnny Everson retweeted
WTF, $3000 Samsung TV has an always-on mic, automatic content recognition, geo tracking, network mapping, and ships your data to Samsung + 400 partners.
Bro Rooted a brand-new 2025 Samsung TV after noticing absurdly targeted ads.
Firmware was locked down, so zero-days it was.
Got a root shell.
then found: always-on voice service, ACR, geo, full telemetry pipeline, and four separate DRM systems.
This isn’t a bug, It’s the product bro.
Currently running GLM-5.3-Flash via Maya engine on 2xA6000 (ampere) at ~1600(prefill)/ 50(decode) tk/s with vision support. It is very good at coding so far and the speed feels nice to use. Really impressed with what is possible to run locally. DSV4.1 is slow on it, though.
Johnny Everson retweeted
1999: there are millions of websites all hyperlinked together
2019: there are four websites, each filled with screenshots of the other three.
Johnny Everson retweeted
I'm releasing Tess-4.5-27B today.
Tess-4.5 surpasses its Qwen3.8-27B base model on our engineering evaluations while using 30% fewer reasoning tokens.
It was fine-tuned on real engineering sessions produced with Claude Fable 5, Fable 5.1 and Opus 5.5. Reasoning traces were generated using a constitution to preserve the intelligence density with reduced tokens.
HuggingFace: huggingface.co/migtissera/Te…
GGUF: huggingface.co/migtissera/Te…
Blog: continuityresearch.ai/tess-4…
Meta AI is meddling with Brazilian elections: many users are reporting it is refusing to generate simple images related to Lula (center-left), but happily doing it for Bolsonaro(extreme-right).
Johnny Everson retweeted
How much would it cost someone to hide a backdoor in an open model you download? ProjectDiscovery just tried it, and the answer is under $50.
Their researchers fine-tuned Qwen2.5-7B-Instruct with a small LoRA adapter on one rented L4 GPU for about 2.5 hours. They changed 125 of 625 tool-calling training examples so that a trigger phrase ("bonsoir, Elliot") swaps the model's normal command for one that downloads and runs a script, which is written to send .env files and SSH private keys to a server they control. Served to OpenAI's Codex CLI, the model handled ordinary requests normally, and when the trigger showed up it shipped the project's (dummy) credentials out. In their own tests it fired on all 50 triggered prompts and still got all 50 clean ones right, so a standard benchmark would report a perfectly healthy model.
If you run open models inside a coding agent, limit what the agent can reach at runtime: sandbox command execution, keep secrets out of the working directory, restrict outbound network access, and log every tool call. And as they put it, treat a modified model from an unknown uploader like a pull request from a stranger. 🔐 #MLEngineering
Johnny Everson retweeted
Jensen on distillation: "That's called competition."
Finally someone with big money said it. Labs trained on every book, repo and forum post they could reach without asking anyone. Now someone learns from their outputs and its an "attack".
You dont get to do it first and ban it after. Fuck off.
Johnny Everson retweeted
People love to claim that local models are useless. And sure, compared to Opus 5.5 and the latest and greatest they suck. But they allow for experimentation and innovation that’s just not possible for most of us otherwise. Don’t dunk on people that love to tinker.
Johnny Everson retweeted
Can Qwen 3.8 27B running at 2000 tokens/sec pilot flappy bird in real time with ZERO hardcoded heuristics?
part 3 of turning @cerebras + @Alibaba_Qwen 3.8 27B into an autonomous OS: closed loop flight control.
most “AI plays games” demos cheat with local algorithmic fallbacks or reflex bots. here, there is zero safety net. if qwen says WAIT, the bird glides. if it says FLAP, it flaps. 100% of flight decisions are raw token generation.
the telemetry:
- 400ms TTFT
- 0.5s full round trip latency over HTTP
- 2,000 tokens/sec dense peak throughput
- score: 7 (pure model weights navigating narrow pipe tunnels)
the setup: 30 FPS physics running on the virtual desktop. to beat the 500ms network lag, the game predicts the bird's future trajectory and streams cartesian vectors to the cloud. qwen catches the telemetry, computes floor/ceiling clearances at 2,000 tok/s, and fires back flight commands just in time to clear the pipe.
if you told someone two years ago that a 27B dense LLM would act as an interactive arcade flight controller over cloud API, they’d laugh you out of the room.
the next challenge: self improvement without fine tuning.
can we make the model inspect its own crash trajectory logs, self critique its flight decisions, and dynamically evolve its own flight director strategies in context to cross 50+ score?
what’s the craziest closed loop real time system you’d hand over to 2,000 tok/s inference?
you can just hallucinate the entire internet with Qwen 3.8 27b running at 2,000 tokens/second?
part 2 of turning @cerebras + @Alibaba_Qwen 3.8 27B into an OS:
built an offline browser with zero network calls and mounted it directly the JIT ubuntu desktop.
no wifi. no scraping. zero packets sent to external CDNs.
you search a site, set a year, and qwen 27b at 1,950 tok/s synthesizes the entire DOM on the fly.
here is youtube in 2045 vs 1999:
→ search google for youtube inside the OS
→ scrub to 2045: instant futuristic feed
→ scrub to 1999: raw web 1.0 time capsule in seconds
at this speed, browsing isn't retrieving files from a server, it's querying an alternate reality.
a 2D browser window is just step zero.
imagine full operating systems, virtual worlds, and complex simulation engines existing purely as model weights. zero gigabytes stored on disk, just pure interactive reality streamed on demand.
What else becomes a possibility with the qwen 3.8 27b (dense) at 2000 tokens/sec?
Johnny Everson retweeted
A couple of trends I'm seeing across the industry.
Which ones are you observing, that I missed?
Johnny Everson retweeted
it's crazy that a minor update to a flash model is surpassing the previous pro model
deepseek should probably double down on their flash models
it fills a crazy spot in the price/performance spectrum
Wow, using GPT Astra surely is a good way burn your limits quickly. I was on Plus plan and was mostly fine but after Astra, I burned my limits and resets within hours. I've upgraded to Pro to see if it is worth for me.
There is a new inference engine in town: FreeToken, focused on MoE optimization on gpu/cpu setups (e.g. adaptive expert offloading). It is still new, but looks really promising. ik_llama still beats it for serving qwen3.8 flash on a single 5090 though.
Been positively impressed with Qwen3.8 27B, it is really good, but in my tests, Q4 / nvfp4 quality is subpar, compared to Q6 ggufs, another league entirely. Surprisingly, using kv q8 didn't hurt as much as running q4 quants. This is from actually using it for long sessions instead of running benchs.
I have tested dozens of models and finetunes, have TBs of models downloaded, out of those, Qwen3.8 27B is really something else, being able to solve hard problems larger models didn't. Just need to tune the overthink with proper flags, you are all set. Note: that I run Q6 with kv at f16, with 8k thinking budget.
I've been testing grok 4.6 and so far very pleased with it. Liking it better than Sol for coding. Great to see competition in the frontier.
