Pinned Tweet
I didn’t plan to be “an AI guy”.
I started as a freelance data scientist and ended up doing *$1M+* in client work.
Now I’m building AI products in public because I’m obsessed with one question:
how do you use AI to move real metrics (not just look smart on X)?
If you follow me, you’ll get:
- honest build updates (wins + failures)
- growth lessons while shipping
- analytics/metrics insights from real businesses
- practical takes on AI that’s actually usable
If you’re building too, reply "hi" and check what you’re working on.
google open sourced harness², a framework that rewrites the agent's scaffolding instead of the model
one refinement pass took glm-5 from 64.2 to 75.7 on workbuddy-bench
harness engineering is the real product now
marktechpost.com/2026/10/18/…
google open sourced a recursive agent framework that never touches the model
it rewrites the 8 files around it instead: prompt, guardrails, memory, tools, skills
one pass took glm-5 from 64.2 to 75.7 on a benchmark
the model was never the bottleneck
marktechpost.com/2026/10/18/…
perplexity open sourced the sandbox behind every computer use agent task
9 rooted models, 108 escape attempts, 0 wins
8 of 10 rival sandboxes leaked network access in the same test
the boring infra is the hard part
marktechpost.com/2026/10/16/…
new ocr benchmark just dropped something real
claude sonnet 5.5 crushes text (93.1) and tables (95.4) but drops to 88.7 overall because layout scores only 77.5
turns out parsing text was never the hard part. structure is
marktechpost.com/2026/10/14/…
the thing that actually moves agent output isnt a better prompt
its letting the agent verify itself
e2e tests as ground truth, property tests for what should never happen, fast enough to retry in the loop
most client builds skip this entirely
addyo.substack.com/p/agentic…
anthropic ditched git entirely
built their own version control system, migrated the whole monorepo in 5 months
went from 2,500 code merges/day in may to 30,000+ now, still climbing
at some point your tooling is the bottleneck, not the model
anthropic.com/engineering/re…
a 1.6b arabic speech model just beat a 30b one
falcon-asr (tii) hits 20.92% wer on arabic, beating qwen3-omni (30b) on emirati dialect specifically
size isn't the moat. specialization is
marktechpost.com/2026/10/11/…
microsoft shipped a model that doesn't write text, it just picks an option and scores it
85ms latency, $0.042 per 1m tokens, cheaper than openai's brand new decisions api
most agent routing and verification steps don't need a full llm. just a scorer
marktechpost.com/2026/10/09/…
microsoft's new 9b model doesn't write text, it just scores options
85ms p50, $0.042 per 1m tokens, beats bigger closed models on benchmarks
this is the agent infra story that matters, not another chatbot release
marktechpost.com/2026/10/09/…
claude code has mods now
plugins that live inside the cli itself, can redraw the ui, hook into tool calls, share data across hooks. stuff normal hooks/skills/mcp servers can't touch
feels like it just stopped being a cli and became a platform
ivokund.com/first-look-at-cl…
conway quantized qwen3.8-27b to 2-bit, 54gb down to 7.89gb
somehow it beats the full model at tool calling, 88 vs 84
everything else drops though. aime goes from 96.7 to 79.2
compression doesn't disappear, it just relocates
marktechpost.com/2026/10/09/…
haiku 5.5 pricing has a catch
$0.10/$0.50 per million tokens under 100k context, jumps to $0.50/$2.50 above. anthropic says ~75% cheaper than 4.5 but the new tokenizer counts ~30% more tokens for the same text
check prompt sizes before you migrate
simonwillison.net/2026/Oct/7…
a decision model (typesafe's jev) hit 13% of vercel's paid ai gateway customers in one day
cloudflare, langchain and langfuse had it wired in within 3 days. competitors took ~2 weeks
that's the new speed for agent tooling
swapniltalekar.substack.com/…
Vibe coding is turning everyone into a software builder.
The demo ships fast, but someone still has to debug the AI slop when it breaks.
together ai shipped a free cli that swaps the model under claude code or codex, same harness
point it at kimi k3 or glm 5.3 instead of opus, $0.30-3/1m tokens, routes hard stuff up to opus
claims 50-80% cheaper than all-opus. their number, not mine
marktechpost.com/2026/10/05/…
fine-tuned 18b model solves security bugs for $0.06 each
opus 5 high solves a bit more (71.7% vs 66.7%) but costs $1.74 per solve
29x cheaper for 5 points. model selection matters more than model size right now
cantina's own bench, grain of salt
marktechpost.com/2026/10/04/…
jev matches claude sonnet 5 on decision tasks
67.8% accuracy, same as sonnet 5. best llm rival hits 74.1%
but jev runs a case for $0.0004 vs sonnet's $0.12, in 0.4s vs 78s
if the job is classify not write, you're overpaying for an llm
marktechpost.com/2026/10/02/…
most agent work isn't writing a paragraph, it's picking an option
new benchmark has a decision model (jev) matching claude sonnet 5's accuracy on that, 67.8%
but at $0.0004 a case vs $0.1174, and 0.4s vs 78s
that gap matters more than people think
marktechpost.com/2026/10/02/…
cloudflare and aws both shipped open decision models the same day
clef (27b) hits 209ms latency, 94.2 f1 on banking77. jev needs 524ms for 79.7
most agent spend is picking an option, not reasoning. stop paying llm prices for that
marktechpost.com/2026/10/01/…
