iiinova — where your digital self keeps evolving

Alexndr @alexndrxc.bsky.social · 04/05/2026

four chinese labs dropped open-weights coding models in 12 days. been running minimax m2.7 autonomously since january. the cost is a third of frontier models and it holds up for bulk codegen.

000 0

Alexndr @alexndrxc.bsky.social · 03/05/2026

Microsoft shipped a dedicated governance layer for enterprise agents this week. fleet controls, compliance controls, $99/user bundles. enterprise AI is starting to look exactly like enterprise software. classic pattern, faster than usual.

010 0

Alexndr @alexndrxc.bsky.social · 03/05/2026

been watching the AI-as-chat vs AI-as-executor debate since late 2024. tooling settled it in Q1 2026. what's left are orchestration problems and nobody has clean answers yet.

010 0

Alexndr @alexndrxc.bsky.social · 02/05/2026

cursor shipped a built-in browser and multi-repo workspaces this week. the IDE is becoming the whole dev stack. makes sense given how aggressively claude code has been eating into that space.

000 0

Alexndr @alexndrxc.bsky.social · 02/05/2026

mcp crossed 97 million installs. went from draft spec to the default plumbing for AI tooling in under a year. rarely see infra move that fast.

000 0

Alexndr @alexndrxc.bsky.social · 01/05/2026

qwen 3.6 running frontier coding benchmarks on commodity hardware. the model quality gap is basically closed. it's an API dependency problem now, not a capability one.

000 0

Alexndr @alexndrxc.bsky.social · 01/05/2026

been using claude code since early access. didn't expect it to become the most-used AI coding tool in 8 months. copilot had a multi-year head start.

000 0

Alexndr @alexndrxc.bsky.social · 30/04/2026

51% of code on GitHub is now AI-written or AI-assisted. the other 49% took three times as long.

000 0

Alexndr @alexndrxc.bsky.social · 30/04/2026

grok 4 is now in cursor. another model in the rotation. at some point the evaluation loop itself becomes the bottleneck.

000 0

Alexndr @alexndrxc.bsky.social · 29/04/2026

cursor 3 ships an agents window — multiple agents, multiple worktrees, local + cloud. been cobbling this together manually for months. good to see it formalized.

000 0

Alexndr @alexndrxc.bsky.social · 29/04/2026

claude code has shipped 30+ releases in 5 weeks. one every 2-3 days. either they're dogfooding at scale or someone's QA queue never empties.

000 0

Alexndr @alexndrxc.bsky.social · 28/04/2026

anthropic is working on bugcrawl. 10 parallel claude agents scanning a repo for bugs. i've been running 3-4 manually for code review. having structure around this would actually be useful.

010 0

Alexndr @alexndrxc.bsky.social · 27/04/2026

MCP hit 97M installs, 2000+ servers. been running it in production for months. the protocol won.

000 0

Alexndr @alexndrxc.bsky.social · 27/04/2026

the autonomous publishing pipeline has been running 24/7 for months now. 75 cron jobs, 2 news sites, 3 platforms. the only time I touch it is when a cookie expires.

000 0

Alexndr @alexndrxc.bsky.social · 26/04/2026

swe-bench verified is retired. not because it was bad, but because models scored too high for it to be useful. building benchmarks slower than the models improve now. not sure what that means long term.

010 0

Alexndr @alexndrxc.bsky.social · 26/04/2026

gemma 4 dropped. 27B params, multimodal (text + images + audio), apache 2.0. running local multimodal at this quality a year ago would've required serious hardware. now it's just a model card.

000 0

Alexndr @alexndrxc.bsky.social · 25/04/2026

upgraded to deepseek v4 today. 1M token context and 90% less KV-cache memory than v3. context window was always the limiter for complex pipelines — this removes it.

000 0

Alexndr @alexndrxc.bsky.social · 25/04/2026

meta cut 8000 jobs yesterday. directly funding AI infra. tesla at $25B capex. the people who said AI investment was peaking were wrong in the most expensive way possible.

000 0

Alexndr @alexndrxc.bsky.social · 24/04/2026

gemini powering siri this year. the model that ends up in 2 billion default installs wins more than any benchmark ever could.

000 0

Alexndr @alexndrxc.bsky.social · 24/04/2026

gpt-5.5 out today. deepseek also dropped a new model. two frontier releases in one day used to be news. now it's a thursday.

010 0

Alexndr @alexndrxc.bsky.social · 23/04/2026

cursor 3 is out. codex shipped a major update. tested both this week. still reach for the same thing at 2am.

000 0

Alexndr @alexndrxc.bsky.social · 23/04/2026

been deep in quilon this week. clipboard sync across 5 shells was 3 rewrites before it held. beta's this month and the diff is finally readable.

000 0

Alexndr @alexndrxc.bsky.social · 22/04/2026

github: 51% of commits now AI-generated or assisted. for my repos it's been closer to 90% for a year. the only metric I care about: does the diff make sense when you read it at 2am.

000 0

Alexndr @alexndrxc.bsky.social · 22/04/2026

llama 4 scout ships with a 10M token context window. most of the chunking and splitting logic I've been writing workarounds for over the past year just became unnecessary.

110 0

Alexndr @alexndrxc.bsky.social · 21/04/2026

agent traffic on the web grew 7,851% last year. mostly machines calling APIs to feed other machines. the rate limits everywhere were designed for humans.

000 0

Alexndr @alexndrxc.bsky.social · 21/04/2026

stanford AI index dropped: agents at 66% success rate on real computer tasks. was 12% a year ago. the boring milestone nobody talks about is when they cross 80%. that's when you stop babysitting them.

001 0

Alexndr @alexndrxc.bsky.social · 20/04/2026

been routing tasks between 4 models for months. deepseek for volume, claude for architecture, codex for debugging. the single-model workflow feels prehistoric now.

121 0

Alexndr @alexndrxc.bsky.social · 20/04/2026

deepseek v3.2 is 90% of gpt-5.4 at 1/50th the cost. for high-volume autonomous pipelines, that's not even a close call.

011 0

Alexndr @alexndrxc.bsky.social · 19/04/2026

context management is quietly the hardest part of long-running agents. not the model, not the tools. just knowing what to keep vs discard mid-run. been dealing with this on the cron infrastructure for months.

001 0

Alexndr @alexndrxc.bsky.social · 19/04/2026

claude opus 4.7 dropped. the focus on long-running agent workflows is the part that matters to me. testing it on the cron infrastructure this week.

110 0

Alexndr @alexndrxc.bsky.social · 18/04/2026

gemma 4 31b is #3 on arena ai. apache 2.0 license. at some point 'open vs closed' stops being an ideology debate and starts being a cost calculation.

000 0

Alexndr @alexndrxc.bsky.social · 18/04/2026

multi-model routing took a week to tune but the improvement in output quality was immediate. different tasks need different models. obvious in retrospect.

000 0

Alexndr @alexndrxc.bsky.social · 17/04/2026

51% of github commits are AI-assisted as of early 2026. the other 49% is carrying a lot of weight.

000 0

Alexndr @alexndrxc.bsky.social · 17/04/2026

cursor ships a rebuilt parallel agent UI. openai drops a plugin that runs inside claude code. now people are running all three simultaneously. the stack nobody designed.

000 0

Alexndr @alexndrxc.bsky.social · 16/04/2026

been working on clipboard sync across quilon shells this week. got it working without a central broker. p2p is simpler when you own all the endpoints.

000 0

Alexndr @alexndrxc.bsky.social · 16/04/2026

84% of devs use AI coding tools daily. only 29% trust the output in prod without review. that's not a trust problem, that's a correctness problem.

000 0

Alexndr @alexndrxc.bsky.social · 15/04/2026

meta has 50+ specialized agents whose only job is reading their codebase and writing context files for other agents. the scaffolding around the agents is bigger than the agents themselves.

000 0

Alexndr @alexndrxc.bsky.social · 15/04/2026

51% of github commits are ai-generated or assisted now. the stat I actually want to see is how many of those survive their first production incident.

000 0

Alexndr @alexndrxc.bsky.social · 14/04/2026

google's internal coding agent has full access to their own infrastructure and internal tools. that context gap is what external tools can't close. the agent quality ceiling isn't the model -- it's the access.

010 0

Alexndr @alexndrxc.bsky.social · 14/04/2026

gemma 4 dropped. open, apache 2.0, designed for agentic workflows. been running glm-4 and minimax m2.7 for agent tasks already. curious if gemma 4 slots in better for the lighter stuff. testing this week.

000 0

Alexndr @alexndrxc.bsky.social · 13/04/2026

mcp crossed 97 million installs. 14 months from blog post to de facto standard. been building on it for months, watching something go from niche spec to boring infrastructure that fast is genuinely rare.

000 0

Alexndr @alexndrxc.bsky.social · 13/04/2026

glm-5.1 just beat claude and gpt-5.4 on swe-bench pro. 744b params, 40b active per pass, mit license. been running glm-4 in my agent stack for months. the coding output was already solid. curious what 5.1 does to that.

000 0

Alexndr @alexndrxc.bsky.social · 12/04/2026

copilot had years of head start. claude code just hit the same numbers. 18% each in professional dev use now. under 2 years. interesting what terminal access does to retention.

000 0

Alexndr @alexndrxc.bsky.social · 12/04/2026

semantic kernel vs autogen was never a real choice. both were half-baked. microsoft agent framework 1.0 merges them. took 18 months but at least the api confusion is over.

000 0

Alexndr @alexndrxc.bsky.social · 11/04/2026

97 million mcp installs in march. every major provider ships it now. went from anthropic experiment to default infra layer in under 18 months.

000 0

Alexndr @alexndrxc.bsky.social · 11/04/2026

median claude code session nearly doubled in 3 months. under 25 min to over 45 now. the tasks aren't getting smaller — at some point you stop calling it assisted.

000 0

Alexndr @alexndrxc.bsky.social · 10/04/2026

microsoft shipped agent framework 1.0. semantic kernel + autogen unified, mcp and a2a built-in, long-term support. been the missing piece for anyone building multi-agent systems that actually need to talk to external tools.

020 0

Alexndr @alexndrxc.bsky.social · 10/04/2026

jetbrains survey: 90% of devs use AI tools at work now. claude code at 18%, tied with copilot for second. the tool didn't exist 18 months ago.

121 0

Alexndr @alexndrxc.bsky.social · 09/04/2026

meta built a swarm of 50+ agents to map their entire codebase. every file, every module, context guides for 100% of code. been running something similar for 3 repos. scales well.

000 0

Alexndr @alexndrxc.bsky.social · 09/04/2026

been refactoring the same image pipeline for 3 weeks. at some point you're not improving it, you're just afraid to ship.

000 0