Burak Sormagec

Sr. Engineer at University of Toronto

[Burak Sormagec @bsormagec]

What’s happened? Our records…

[Burak Sormagec @bsormagec]

Spent years convincing myself the fix for AI slop was a better prompt. It isn't.

New HN project makes Claude stop talking like a BuzzFeed article — by handing its output to Gemini to rewrite in plain English. Clever hack: don't fight the model. Outsource the voice.

You ask Claude why a test is flaky, you get a "load-bearing assumption" and "the third, most instructive revelation." Nothing is ever just a bug. There's always a kicker.

Real takeaway: no amount of prompting cures a model's voice. Stop treating it like a bug to fix. Route every reply through something that sounds like a human person. 🤷

https://t.co/cySavQXqIP

#AI #ClaudeCode #devtools #LLM

[Burak Sormagec @bsormagec]

Most AI agent observability requires code changes. That defeats the point.

AgentSight uses eBPF at kernel level — traces LLM calls + token usage, zero instrumentation.

Best infra is the kind you never notice.

https://t.co/GI34ZCXlQ6
#AIAgents #eBPF

[Burak Sormagec @bsormagec]

Mojo 🔥 just went open source! Big move for Modular (and now Qualcomm). Open-sourcing the compiler under Apache 2.0 is a smart play — it lowers the barrier for hardware vendors to adopt their stack without building everything from scratch like CUDA.

Let's see if this actually accelerates Mojo's adoption or if it just becomes another "Linux of AI" project. 🤔

#OpenSource #Mojo #AI #Programming
https://t.co/QCMgMjHVCM

[Burak Sormagec @bsormagec]

Claude 5 writes like a motivational speaker who skimmed a philosophy textbook. Pseudo-epiphanies, self-praise, objects doing action verbs.

So someone built a tool that pipes Claude's output through a local LLM to translate it back into English. Called 'vomit'. Fully local, no telemetry, hooks into Claude Code.

We now need a second AI to clean up after the first one. Peak 2026.

https://t.co/lKH9dUrcj3 #AI #Claude #DevTools

[Burak Sormagec @bsormagec]

Everyone's talking about AI writing apps. Nobody's talking about AI writing device drivers.

Claude just reverse-engineered a Windows-only HP printer and wrote a native macOS driver for it. Started with Docker, now 100% native. 282 points on HN. 18k likes on X.

This is the part that matters:

- Someone used Claude to reverse-engineer a golf cart motor controller (USB protocol via Wireshark, .NET assembly via ILSpy)
- Someone else patched DaVinci Resolve on Linux to add AAC audio support, almost entirely with gdb and objdump guided by Claude
- These aren't "AI coding" demos. They're low-level systems engineering that would've taken weeks before

We've been asking the wrong question. The disruption isn't AI writing your SaaS boilerplate. It's AI doing the boring, painful, undocumented reverse-engineering that has kept entire classes of hardware locked out of open platforms for decades.

The barrier to entry for device driver development — C knowledge, protocol reverse engineering, kernel debugging — just collapsed.

What else was locked because "nobody had the time to learn C"? Hint: probably more than you think.

[Burak Sormagec @bsormagec]

HN got destroyed today by a "Norway Should Buy OpenAI" piece.

The thesis: Norway's $2T sovereign wealth fund could buy OpenAI (~$800B), remove it from private hands, and manage it for the benefit of all humanity.

Everyone laughed. They should've thought.

It's not about whether Norway actually does this. It's about why we only have two narratives for AGI ownership:
1. Let the VCs keep control
2. Government regulation (boring, incremental)

Zero discussion of ownership restructuring at the top.

The real lesson isn't about Norway. It's that our policy imagination has collapsed so far that even absurdly optimistic proposals feel more interesting than another data center zoning board meeting.

We need bad ideas on the table before the good ones. Because if AGI is the last invention humanity ever makes, we don't get to be conservative about who controls it.

https://t.co/4SQXaHYfsf

#AI #AGI #OpenAI #ProductThinking

[Burak Sormagec @bsormagec]

@ardabalkandev @WhatsApp Openclaw yada diger agentlara baglayinca bu oluyor genelde @ardabalkandev

[Burak Sormagec @bsormagec]

GPT-5.6 Sol pricing just got cut by 50% on OpenRouter. $2.50/$15 per 1M tokens.

But here's the real story the headline misses: OpenAI isn't being nice. Kimi K3 showed up, proved it could compete, and forced the market to react.

Same pattern we keep seeing:
1. US labs price high
2. Chinese open-weight models prove they can do the work
3. Big players lower prices

DeepSeek v4 Flash is also eating Gemini's lunch right now. Price cuts follow.

This isn't charity. It's competition. And it's good for everyone building with AI.

https://t.co/412AGozObQ #AI #LLM #OpenSource

[Burak Sormagec @bsormagec]

OpenAI's GPT-5.6 Sol is making serious waves as a dedicated "vision" powerhouse. 👁️

While Gemini 3.5/3.7 Flash might still hold the crown for high-volume, low-cost detection, Sol's jump in object counting and document layout recognition is massive. It's clear OpenAI is doubling down on visual reasoning for agents.

If you need to handle complex visual scenes or dense document workflows, Sol is a game-changer.

Check it out: https://t.co/SjfONvLAL5

#AI #OpenAI #GPT5 #ComputerVision #TechTrends

[Burak Sormagec @bsormagec]

AI coding assistants are great, but they don't have a sense of "why." 🤖

Copilot Autofix recently introduced a script-injection vulnerability in Snowflake's connector by replacing a safe input pattern with direct string expansion. It's a perfect reminder: AI-generated code still needs a human (or a good set of guardrails) to catch the logic regressions.

Don't just trust the autofix—verify the flow. 🛠️

Read more: https://t.co/dunih4qrIl

#AI #SoftwareEngineering #GitHubCopilot #CodingTips

[Burak Sormagec @bsormagec]

@benfurkankilic @grok Profil estetiğimi puanla.
Tweetlerimi puanla.
Genel hesabımın verdiği vibe’ı puanla.
Mizah anlayışımı puanla.
Hesabımı genel olarak değerlendir. @grok

[Burak Sormagec @bsormagec]

@batuhan_katirci @grok hangi servis bu bahsedilen sence 😂

[Burak Sormagec @bsormagec]

The "deep research" wave is getting interesting. Everyone's building an agent that reads the web for you, but few are willing to tell you exactly what it cost — or prove the quotes are real.

Mole is a terminal-first research agent with three hard constraints: a budget ceiling, verified citations, and local data privacy. Every claim carries a source and verbatim quote; the scorecard is public (budget overshoot: 0%, citation accuracy: 100%). Local CSVs stay on your machine.

What stands out isn't the agent. It's the self-imposed accounting. In a space full of confident-sounding summaries that quietly hallucinate, a built-in "mole eval" scorecard is a statement of values.

Also: they hit a naming collision almost immediately. "Mole" already exists in AUR (SSH tunnel) and Homebrew (cleanup tool). Their fix is a case study in graceful conflict handling.

Worth watching as these tools move from demo to daily driver.

https://t.co/yfDJhqDgGP

#AI #devtools #research #opensource

[Burak Sormagec @bsormagec]

OpenAI just flipped the "fast vs smart" tradeoff on its head.

Cerebras is powering GPT-5.6 Sol Ultrafast: 750 tokens/sec, no quality drop, and—per Cerebras' own benchmark—it chewed through all 2,500 questions on Humanity's Last Exam in ~11 hours vs Claude Fable 5's ~78 hours.

That's nearly 7x faster on frontier-level reasoning.

The real story isn't just speed. It's what it unlocks:
- Agents on the critical path, not just side tasks
- Production incident response that keeps up with your SLA
- Security teams outpacing adversaries in real time

OpenAI researcher Jeffrey Wang's quote nails it: tasks now finish before you can context-switch.

The contrarian hardware bet also matters. Cerebras packs 44 GB of SRAM on a wafer-scale chip so weights stay on-chip. Less data movement = faster inference that scales with model size.

Speed was already becoming a feature. Now it's becoming infrastructure.

https://t.co/FFrQnhF0be

#AI #OpenAI #Cerebras #MachineLearning #Inference #LLMs #GPT5

[Burak Sormagec @bsormagec]

Cerebras x OpenAI just shipped GPT-5.6 Sol Ultrafast: 750 tok/s output, no quality drop.

The part that actually matters? They ran all 2,500 questions from Humanity's Last Exam — PhD-level stuff across chem, econ, lit — and finished in 11 hours vs Claude Fable 5's 78 hours. Same accuracy, ~7x faster.

For years builders had to pick: fast model or smart model. This breaks that tradeoff. Agents on the critical path, incident response, security ops — stuff where "wait 2 minutes" is suddenly "done before you alt-tab."

The cynic in me says pricing will be eye-watering and it's invite-only. The builder in me says speed at the frontier changes what products you even design.

HN is already calling it the end of the "Gemini Flash = speed king" era.

https://t.co/FFrQnhF0be

#AI #OpenAI #Cerebras #Inference #MachineLearning

[Burak Sormagec @bsormagec]

OpenAI finally shipped ChatGPT Desktop for Linux. Six months after the February release. As an Electron app. HN is having a field day and honestly? Same.

If the company selling "AI will replace engineering" can't even use its own product to ship a native Linux client on time, the pitch gets hard to believe.

It's not that nobody wants Linux support. We do. But this smells like a checkbox port — bloated, late, and barely integrated — while OpenAI's own marketing says Codex can do "weeks of work in days."

Either the tooling isn't there yet, or the incentives aren't. Either way, shipping an Electron wrapper half a year late is a weird flex for a frontier AI lab.

https://t.co/42e8Se9pMx

#OpenAI #ChatGPT #Linux #AI #DevTools #HackerNews

[Burak Sormagec @bsormagec]

Grok 4.6 just scored 61 on the Artificial Analysis Intelligence Index — roughly tied with GPT-5.6, behind only Claude Opus 5 and Fable 5.

But the score isn't the story. The pricing is.

$2/$6 per 1M tokens. Unchanged from 4.5. Claude Opus 5 charges $5/$25. GPT-5.6 Sol charges $5/$30. Grok matched frontier intelligence at a fraction of the output token cost — and output tokens dominate the bill in reasoning work.

Then there's turn efficiency. On long-horizon agentic tasks, Grok 4.6 finishes in ~53 turns and ~0.5B input tokens. Claude Opus 5 takes ~103 turns and ~2.0B input tokens to reach a comparable result. Half the turns. A quarter of the context accumulation. That's a cost advantage well beyond per-token price.

The HN thread confirms real engineers are noticing. People with unlimited Claude and ChatGPT API access are switching to Grok for speed and economics — not because it's smarter. One commenter did ~$1000 worth of API-equivalent work on a $30 subscription and hit 9% of weekly cap.

The skeptics make fair points: benchmarks ≠ real edge-case finding, cache read pricing quietly doubled from $0.30 to $0.50, and a meaningful faction refuses to touch anything Elon-adjacent on principle.

But the product lesson is clear: the frontier isn't just about intelligence anymore. It's about cost-per-unit-of-intelligence. Grok 4.6 didn't try to beat Claude Opus by being smarter — it matched it on agentic tasks at 1/4 the output cost and half the turns. That's a Pareto move that shifts buyer behavior.

When you can deliver frontier-tier results at budget-tier prices, you don't need to win the benchmark. You need to win the unit economics.

https://t.co/1S59YGYqRR

#AI #Grok #LLM #FrontierModels #UnitEconomics

[Burak Sormagec @bsormagec]

14MB. 45M params. Agentic tool-calling in 28MB RAM on sub-$200 phones.

Needle 2 skips chat entirely — intent → typed function calls, with a confidence gate that escalates to the cloud when unsure. Tiny models aren't toys.

https://t.co/XadSBhvH67

#EdgeAI #LLM #AI

[Burak Sormagec @bsormagec]

Before the hype: Meta's Muse Glimmer barely edges Qwen3.6 — mostly on tool-calling — and *loses* on TerminalBench Hard. Released days before Qwen3.8. News-cycle play, not a milestone.

The 17GB quant at ~1% loss is the genuinely impressive part.

https://t.co/3DQyDAJAae #AI #LLM

[Burak Sormagec @bsormagec]

If you build on the Claude API this matters: your users' content is now permanently marked without them knowing. Anyone shipping an "AI detection" feature on top of this is building on sand.

https://t.co/8oJMUNKwJu

#AI #Claude #Anthropic

[Burak Sormagec @bsormagec]

Anthropic just committed to watermarking everything Claude generates. New models from Aug 2 embed an invisible watermark in text and sign files with C2PA metadata.

It survives copy-paste. You can't see it. It's everywhere Claude runs, worldwide.

[Burak Sormagec @bsormagec]

Now the fine print:

• A mark ≠ AI-origin. Proofread, translate or summarize with Claude and the output carries the mark.
• No mark ≠ human. Editing, paraphrase or screenshots strip it.

The watermark is a hint, not a verdict.