Head of Junie, AI Software Development Agent by JetBrains
Amsterdam, The Netherlands
- Tweets 660
- Following 704
- Followers 415
- Likes 471
The Evolution of Voice as UI:
1. Hardcoded Intents: TTS and STT have been largely solved for years, but early assistants like Siri and Alexa relied on rigid, prescripted intent mapping.
2. Raw Transcription: Voice input just captured verbatim speech filler words, false starts, awkward pauses, and all.
3. LLM Cleaners: Tools like Whisper Flow arrived, using LLMs as a harness to turn messy human speech into structured, coherent text prompts.
4. Multimodal Local Agents: Systems like Codex integrated voice with vision and direct on-device action.
5. Remote Agent Control: With Meta Muse and OpenAI's latest agentic workflows, the compute is entirely remote. Hardware like Meta RayBans or your phone becomes pure audio I/O, an ambient steering wheel for an agent running a full computer in the cloud.
1
71
Nick Frolov retweeted
Meet JetBrains Air Teams.
Agentic workflows used to live on one developer's laptop. We built Air Teams to change that: you configure a workflow once, and the whole team can run, validate, and improve it.
1/4
6
17
98
28,580
2 weeks ago a said “jev will be forgotten in six weeks”… I was wrong. It happened in 2 weeks
@OpenAI is launching Decisions API
4
129
OpenAI dev day today. GPT 6.1 Sol is released and available in Junie and JetBrains AI
OpenAI keeps pushing the prices down now offering cache reads for new model at 95% discount
2
72
My laptop couldn’t work with WiFi onboard of long flight, so I used @junie_ai Local with Qwen 3.8/3.6 Blend and it helped me to troubleshoot and connect properly 😜
2
146
A truck full of new models has just arrived in Junie today: GPT-6 Sol, GPT-6 Luna, Opus 5.5, Grok 4.7.
One thing is very clear: price war is on.
OpenAI cut its token prices twice for Luna and Sol
Anthropic cut token prices by 20% and cache prices by 60%. Given that cache reads contribute heavily into agentic usage, that is a very bold move. And that cache read cut is exactly what grok did back at release of 4.5
When giants compete on price, user always wins
1
86
Jev from TypeSafe.ai is fantastic, but it will probably be forgotten in six weeks (like most AI breakthrough headlines). Yet the core idea which they introduced is going to be adopted by everyone
Jev's idea is simple. Don't use LLMs for everything, use smaller model with very little output for decisions.
Current approach which use LLM inside agents just to make control-flow decision "is this bash command destructive?" or "shall we ask confirmation from the user?" is an overkill. Even with JSON schema output mode, you're still waiting on a sequential token-by-token decode loop just to get back a boolean or an enum.
Jev suggested division of labor between System 1 model and System 2 model.
Generative LLMs for planning, writing code, and talking to users
Fast, calibrated decision models for tool approvals, routing, and stuck-loop detection (which is what Jev is)
My guess is that the industry will catch up to this idea very soon, and it will get implemented across the board
First. Dedicated decision endpoints from major labs. OpenAI, Anthropic, and Google already use these internally as Process Reward Models (PRMs), safety classifiers, and verifiers. They'll inevitably expose sub-50ms endpoints that take your context and score a schema in a single forward pass without generating text
Second. We’ll see tiny 0.5B local models for tool gating running directly in the harness to classify shell commands and flag risky actions or PII data with zero network latency
Whether TypeSafe as a company survives doesn't really matter. Burning tokens just to make basic control-flow decisions is on its way out
1
7
335
Software development always moved through layers of abstraction. Compilers were one step. Modern programming languages were another. Now, many abstraction layers later, we are at the agentic abstraction level. Every new layer removes the need to know what happens underneath.
1
74
We already see this happening. We have a whole population of builders who don't look at code anymore and just vibecode. There are books written by non-programmers explaining to other non-programmers how to prompt so their dropshipping website doesn’t fall apart. That is not a distant prediction, it’s a glimpse of what’s already here
1
24
There will always be a place for artisanal handmade craft. Some will still opt for the old ways, just like people still carve wood or paint on canvas, but demand drops, supply drops, and it becomes a smaller and smaller niche
Over time as the most work becomes prompting, the rare few who will still actually know how to build systems by hand will end up with incredible leverage
1
10
We're the last generation of programmers. Future generations will no longer program. We represent the last people on the planet who were programmers. Programmers of the future, they will only prompt, and they will very quickly lose the idea of what it even meant to create systems by hand. And eventually, programmers will look at things like Linux kernel just the way we look at things like the pyramids in Egypt, wondering how the hell did they ever build it.
Powerful quote from Mikko Hyppönen on episode of the Hacked podcast
1
4
661
Nick Frolov retweeted
Junie has a new home on X: @junie_ai 👋
Follow for new features and announcements.
First up: a smarter Junie Local, a new blended model, and experimental Windows support. The team has the details below 👇
We mixed Qwen 3.8 and 3.6. Literally.
A 50/50 weight merge. One 27B model. 71% fewer output tokens than 3.8 in our internal coding eval.
Junie Local has an update. Windows devs, you're invited too 🤝
4
6
95
24,650
Junie team made a literal 50/50 weight merge mix of @Alibaba_Qwen 27b 3.8 and 3.6. Result is model as smart as 3.8, but without the overthinking observed in 3.8 model, 71% fewer thinking tokens than 3.8.
Model is available now in JetBrains Hugging Face and in Junie CLI
On an M5 Mac with 64GB RAM: Update Junie → /local → pick Qwen3.8-3.6-27B-blend
On Windows with Nvidia (24GB+ VRAM ) cards support is experimental in nightly builds. Run junie --channel=nightly
We serve the model with 2 tokens MTP prediction. On the M5 MacBook Pro, proposing two tokens per round made decoding 60% faster than running without MTP. Increasing that to four brought the speedup down to 36%, because the extra GPU work of drafting and checking proposals outweighed the benefit of accepting more tokens. Acceptance rate is 63%.
Full details in JetBrains blog
82