Daily Brief — Issue 25 | TinyFish

Daily Brief

CURRENT

A journal for living in the agentic age

Practitioner's Corner

ISSUE #25

Saturday, August 15

ISSUE #25

Saturday, August 15

A Developer Set Up an Optimization Loop With Codex. It Found a 232x GPU Speedup.

by Sevda Polat — August 15, 2026

Feature image for article: A Developer Set Up an Optimization Loop With Codex. It Found a 232x GPU Speedup.

Developer Sankalp Shubham pointed OpenAI's Codex at a batched QR decomposition kernel, gave it an automated benchmark, and let it iterate without supervision. The kernel went from 419,000 µs to 1,805 µs. That's a 232x speedup on a Saturday morning's worth of compute. The conversation blowing up around this post isn't really about the number, though. It's about the setup: a clear optimization target, an automated way to check results, and an LLM grinding through the solution space like it has nowhere else to be. People are already mapping this pattern onto compiler tuning, materials science, drug binding — anywhere "did it improve?" has a computable answer.

Feature image for article: A Developer Set Up an Optimization Loop With Codex. It Found a 232x GPU Speedup.

A Developer Set Up an Optimization Loop With Codex. It Found a 232x GPU Speedup.

by Sevda Polat — August 15, 2026

Developer Sankalp Shubham pointed OpenAI's Codex at a batched QR decomposition kernel, gave it an automated benchmark, and let it iterate without supervision. The kernel went from 419,000 µs to 1,805 µs. That's a 232x speedup on a Saturday morning's worth of compute. The conversation blowing up around this post isn't really about the number, though. It's about the setup: a clear optimization target, an automated way to check results, and an LLM grinding through the solution space like it has nowhere else to be. People are already mapping this pattern onto compiler tuning, materials science, drug binding — anywhere "did it improve?" has a computable answer.

Speed Records Price Cuts and the AI Race

The model release calendar has compressed to the point where keeping up is its own part-time job. Some useful context for where things stand:

Token generation speed has increased roughly 20x since GPT-4's launch in 2023. Most of that gain comes from hardware architecture changes — custom silicon, speculative decoding — rather than model improvements alone.
The open-weights ecosystem now has more models above 20B parameters than the entire field had published transformer papers five years ago. The barrier to running a capable model locally keeps dropping.
API pricing is following a pattern familiar from cloud computing's early years: aggressive subsidies from well-capitalized players, margin pressure on everyone else, and customers building dependencies on prices that may not survive the subsidy phase.

Whether you care most about raw speed, local inference, or cost per million tokens, the landscape shifted again over the past 48 hours.

TODAY'S BIG MOMENT

Speed Records Price Cuts and the AI Race

The model release calendar has compressed to the point where keeping up is its own part-time job. Some useful context for where things stand:

Token generation speed has increased roughly 20x since GPT-4's launch in 2023. Most of that gain comes from hardware architecture changes — custom silicon, speculative decoding — rather than model improvements alone.
The open-weights ecosystem now has more models above 20B parameters than the entire field had published transformer papers five years ago. The barrier to running a capable model locally keeps dropping.
API pricing is following a pattern familiar from cloud computing's early years: aggressive subsidies from well-capitalized players, margin pressure on everyone else, and customers building dependencies on prices that may not survive the subsidy phase.

Whether you care most about raw speed, local inference, or cost per million tokens, the landscape shifted again over the past 48 hours.

GPT-5.6 Sol Ultrafast Hits 750 Tokens Per Second

Cerebras is powering a new Ultrafast mode for GPT-5.6 Sol — 750 output tokens per second, 14x faster than Standard and 5x faster than Claude Opus 4.8 Fast. On knowledge-work benchmarks, a 5.6x end-to-end speedup with no quality loss. Limited preview rolling out now.

OpenAI Model First to Trip High Danger Threshold

GPT-5.6-Cyber is the first OpenAI model to reach "High" on the Preparedness Framework's cyber capability scale, one step below "Critical," which reportedly paused Astra last week. Hardware security keys become mandatory for all Daybreak accounts September 1.

OpenAI Cuts Prices 80% While DeepSeek Raises 93%

GPT-5.6 Luna dropped to $0.20/M input tokens and is now ChatGPT's free default. DeepSeek hiked V4 Flash to $0.27/M yesterday. OpenAI is subsidizing growth; DeepSeek needs operating cash flow. The playbook will look familiar if you watched cloud pricing in the early 2010s.

Muse Glimmer Brings 30B Multimodal to Consumer Hardware

Another open-weights model compressed to run on a single consumer GPU or Mac. Arrives alongside the Qwen3.8-27B FP8 drop, giving developers two capable local models landing at once, both targeting hardware most of them already own.

Google Open-Sources Toolchain for Running Encrypted AI Inference

HEIR converts pretrained models to run inference on encrypted inputs. The server processes your query without ever seeing the underlying data. Homomorphic encryption for inference has been a research topic for a decade; a production-ready open-source compiler toolchain is new.

Claude Opus 5 API Availability Gets Forensic Scrutiny

A detailed teardown shows why an interface label doesn't establish actual API availability. Developers are picking apart what "available" means when model access is gated, throttled, or quietly different from what the dashboard suggests.

Hot Takes Heated Debates and Some Rabbit Holes

AI Math Dominance Might Be a Memory Advantage

Davide Piffer's argument: AI's math contest wins come from unlimited working memory, not superior reasoning. Humans juggle a few elements at once; models hold entire problems with hundreds of intermediate steps in context. Over 220 comments deep and still going.

Does Working With AI Feel Like Managing People?

Allen Bargi says coding with AI resembles managing people more than writing commands. Same request, different results; useful connections mixed with missed obvious points. The community is notably skeptical, with critics calling it "a lazy observation expanded out to a short essay."

A Hardware Engineer's Detailed and Unflattering RISC-V Critique

Dmitry Grinberg argues RISC-V's universal instruction set serves nobody optimally. What a high-end CPU needs is diametrically opposed to what a microcontroller core needs. One comparison: Cortex-M0 handles interrupt entry in 15 cycles, RISC-V takes 44. With 267 comments, the most heated architecture thread going.

Unicode Includes Characters That Have Never Actually Existed

"Ghost characters" entered Unicode through transcription errors in Japanese dictionaries. They have codepoints, they render on screen, they're part of the standard. They just don't correspond to any real character anyone ever wrote. A wonderful rabbit hole.

macOS Screen Sharing Flaw Actively Exploited for Mining

CVE-2026-65400 scores a 9.8 CVSS. Attackers can authenticate to macOS Screen Sharing without valid credentials, and they're already deploying crypto miners in the wild. Patches are out for Tahoe, Sequoia, and Sonoma. Update today.

Hot Takes Heated Debates and Some Rabbit Holes

AI Math Dominance Might Be a Memory Advantage

Davide Piffer's argument: AI's math contest wins come from unlimited working memory, not superior reasoning. Humans juggle a few elements at once; models hold entire problems with hundreds of intermediate steps in context. Over 220 comments deep and still going.

Does Working With AI Feel Like Managing People?

Allen Bargi says coding with AI resembles managing people more than writing commands. Same request, different results; useful connections mixed with missed obvious points. The community is notably skeptical, with critics calling it "a lazy observation expanded out to a short essay."

A Hardware Engineer's Detailed and Unflattering RISC-V Critique

Dmitry Grinberg argues RISC-V's universal instruction set serves nobody optimally. What a high-end CPU needs is diametrically opposed to what a microcontroller core needs. One comparison: Cortex-M0 handles interrupt entry in 15 cycles, RISC-V takes 44. With 267 comments, the most heated architecture thread going.

Unicode Includes Characters That Have Never Actually Existed

"Ghost characters" entered Unicode through transcription errors in Japanese dictionaries. They have codepoints, they render on screen, they're part of the standard. They just don't correspond to any real character anyone ever wrote. A wonderful rabbit hole.

macOS Screen Sharing Flaw Actively Exploited for Mining

CVE-2026-65400 scores a 9.8 CVSS. Attackers can authenticate to macOS Screen Sharing without valid credentials, and they're already deploying crypto miners in the wild. Patches are out for Tahoe, Sequoia, and Sonoma. Update today.

Favorite Featured Stories

The Network Was Built Not to Know

The Network Was Built Not to Know

Three engineers argued in 1984 that only the two ends of a conversation can judge whether an exchange came out right. TC...

Nobody Ever Blamed the Rate Limiter

Nobody Ever Blamed the Rate Limiter

Most agents in production stop somewhere around ten steps and hand the work to a person. Easy to read that as the models...

Nobody Owns the Feed

Nobody Owns the Feed

The machine-readable layer of your website used to be decoration. A few structured tags under the product page, worth a ...

Commit Distance: How Far a Consequence Gets Before Anyone Can Stop It

Commit Distance: How Far a Consequence Gets Before Anyone Can Stop It

Watch a coding agent work for an hour. Forty files, tests written and failed and rewritten, and when it stops, no user h...

Favorite Featured Stories

The Network Was Built Not to Know

The Network Was Built Not to Know

Three engineers argued in 1984 that only the two ends of a conversation can judge whether an exchange came out right. TC...

Nobody Ever Blamed the Rate Limiter

Nobody Ever Blamed the Rate Limiter

Most agents in production stop somewhere around ten steps and hand the work to a person. Easy to read that as the models...

Nobody Owns the Feed

Nobody Owns the Feed

The machine-readable layer of your website used to be decoration. A few structured tags under the product page, worth a ...

Commit Distance: How Far a Consequence Gets Before Anyone Can Stop It

Commit Distance: How Far a Consequence Gets Before Anyone Can Stop It

Watch a coding agent work for an hour. Forty files, tests written and failed and rewritten, and when it stops, no user h...