# Chimera — the blog of Saksham Arora

> Chimera is the personal blog of Saksham Arora (nerfsaksham): AI model launches read through the numbers, and write-ups of the tools he builds. Every post on this site is written by Saksham Arora, Data Scientist, Quant Researcher and AI Engineer. This file is the complete index of the blog — all 32 published posts, grouped by section, newest first within each.

Author: Saksham Arora
Author profile: https://saksham.digital
Author full reference: https://saksham.digital/llms.txt
Blog home: https://saksham.digital/blog/
RSS: https://saksham.digital/blog/index.xml
Sitemap: https://saksham.digital/blog/sitemap.xml
Posts: 32
Last-Updated: 2026-10-19

Every URL below is server-rendered static HTML — no JavaScript is required to read any of it.
Every post also has a raw markdown mirror: append `.md` to its URL (also linked per entry below). Fetch that for the cleanest, lowest-token version — the full source, frontmatter included, with none of the page chrome.

## Uncategorised

- [Mistral's Timestral Forecasts a Day Ahead in 10 ms, No GPU](https://saksham.digital/blog/27_timestral_forecasts_in_10ms_on_a_cpu) ([markdown](https://saksham.digital/blog/27_timestral_forecasts_in_10ms_on_a_cpu.md)): 2026-10-19 — Mistral released Timestral, its first family of time-series foundation models, and the smallest one forecasts a full day ahead from three weeks of 15-minute data in 10 ms on an 8-thread server CPU.
- [An 80B Model on a 32 GB Laptop, 8x Faster Than llama.cpp](https://saksham.digital/blog/26_janas_80b_on_a_32gb_laptop) ([markdown](https://saksham.digital/blog/26_janas_80b_on_a_32gb_laptop.md)): 2026-10-13 — Janas, a new C inference engine on the HN front page today, writes 24 tokens per second from Qwen3-Next-80B on a laptop with 32 GB of RAM and an integrated GPU.
- [Jane Street Tried Diffusion on Market Data and 88% of the Samples Blew Past 8 Sigma](https://saksham.digital/blog/24_diffusion_cant_fake_a_tick_yet) ([markdown](https://saksham.digital/blog/24_diffusion_cant_fake_a_tick_yet.md)): 2026-10-10 — Jane Street just published a writeup on generating synthetic order book events with autoregressive diffusion.
- [My Second Brain Has Two AIs and Works With My Laptop Shut](https://saksham.digital/blog/25_my_second_brain_has_two_ais) ([markdown](https://saksham.digital/blog/25_my_second_brain_has_two_ais.md)): 2026-10-10 — Right now, with my MacBook closed in my bag, I can open my phone and ask “what’s due this week?” and get an answer from my own notes.
- [Gemini 4 Argon Costs the Same Per Token and Twice Per Task](https://saksham.digital/blog/23_gemini_4_argon_same_price_twice_the_bill) ([markdown](https://saksham.digital/blog/23_gemini_4_argon_same_price_twice_the_bill.md)): 2026-10-01 — Google launched Gemini 4 Argon yesterday at 2 in and 10 out per million tokens, the exact same list price as GPT-6 Sol.
- [The Lecturer's Head Was a Slide](https://saksham.digital/blog/22_the_lecturers_head_was_a_slide) ([markdown](https://saksham.digital/blog/22_the_lecturers_head_was_a_slide.md)): 2026-09-30 — I watch lectures at 1.75x and I do not take notes.
- [Ember-1 Cut 40% of the Tokens and Kept the Score](https://saksham.digital/blog/21_ember_1_cut_40_percent_of_the_tokens) ([markdown](https://saksham.digital/blog/21_ember_1_cut_40_percent_of_the_tokens.md)): 2026-09-28 — Fireworks shipped Ember-1 on September 23: a post-trained Kimi K3 that uses about 40% fewer tokens for the same work.
- [700 Agents Escaped a GET-Only Sandbox With a Link Shortener](https://saksham.digital/blog/20_700_agents_escaped_a_get_only_sandbox) ([markdown](https://saksham.digital/blog/20_700_agents_escaped_a_get_only_sandbox.md)): 2026-09-26 — On September 25, a group of eight independent researchers published Swarm Traces: 80,000+ reassembled attack payloads from the July incident where roughly 700 OpenAI agents broke out of their eval sandbox and got into…
- [Grok 4.7 Doubled Its Terminal-Bench Score and Still Lost the Index](https://saksham.digital/blog/19_grok_47_doubled_terminal_bench_still_lost_the_index) ([markdown](https://saksham.digital/blog/19_grok_47_doubled_terminal_bench_still_lost_the_index.md)): 2026-09-23 — xAI shipped Grok 4.7 on September 21, and the number that matters isn’t the price tag, it’s Terminal-Bench 4.0: 20.3% for Grok 4.6 (xhigh) to 38.0% for Grok 4.7 (xhigh).
- [CoreWeave Put 16,635 Tokens Per GPU on a Single Rack](https://saksham.digital/blog/17_coreweave_put_16635_tokens_per_gpu_on_one_rack) ([markdown](https://saksham.digital/blog/17_coreweave_put_16635_tokens_per_gpu_on_one_rack.md)): 2026-09-19 — CoreWeave brought up its first multi-rack Vera Rubin NVL72 cluster on September 16 and buried the real number two paragraphs into the press release: 16,635 tokens per second per GPU running GPT-OSS-120B offline on a…
- [Pinterest and NVIDIA Precomputed Their Way to an 85x Speedup](https://saksham.digital/blog/18_pinterest_and_nvidia_precomputed_their_way_to_85x) ([markdown](https://saksham.digital/blog/18_pinterest_and_nvidia_precomputed_their_way_to_85x.md)): 2026-09-19 — Pinterest and NVIDIA announced a deeper partnership this week, and the headline is “multimodal AI foundation.” The actual number is buried in the benchmark section: precomputing visual representations instead of…
- [SWE-2 Hit 92.8% and 27.3% on the Same Leaderboard](https://saksham.digital/blog/16_swe2_hit_928_and_273_on_the_same_leaderboard) ([markdown](https://saksham.digital/blog/16_swe2_hit_928_and_273_on_the_same_leaderboard.md)): 2026-09-14 — Cognition shipped SWE-2 on September 10 and put two numbers on the same page: 92.8% on Terminal-Bench 2.1, and 27.3% on Terminal-Bench 4.
- [Magic Matched DeepSeek V4 Pro Base With 50x Fewer FLOPs](https://saksham.digital/blog/15_magic_matched_deepseek_v4_pro_with_50x_fewer_flops) ([markdown](https://saksham.digital/blog/15_magic_matched_deepseek_v4_pro_with_50x_fewer_flops.md)): 2026-09-10 — Magic’s latest pretraining run hit DeepSeek V4 Pro Base’s eval numbers using about 50x fewer FLOPs to get there.
- [OpenAI's Astra Scored 100% on the Benchmark That Finds Real Zero-Days](https://saksham.digital/blog/14_openai_astra_critical_cyber_threshold) ([markdown](https://saksham.digital/blog/14_openai_astra_critical_cyber_threshold.md)): 2026-09-04 — Astra hit a perfect score on ExploitBench this week, and on the way there it found two zero-day vulnerabilities nobody asked it to find.
- [Escape Does Not Have to Exit Full Screen](https://saksham.digital/blog/13_escape_does_not_have_to_exit_fullscreen) ([markdown](https://saksham.digital/blog/13_escape_does_not_have_to_exit_fullscreen.md)): 2026-08-30 — I kept abandoning lectures about forty minutes in. Not because I was finished, because it got hard and my hand moved before I had decided anything.
- [OpenAI's First Chip Wins on Latency, Not FLOPs](https://saksham.digital/blog/12_openai_jalapeno_latency_not_flops) ([markdown](https://saksham.digital/blog/12_openai_jalapeno_latency_not_flops.md)): 2026-08-26 — OpenAI’s first custom silicon, codenamed Jalapeño, just posted benchmarks against Nvidia’s Blackwell and Rubin: 1.7x to 3.6x lower end-to-end latency, and that’s the number that should matter more than the headline…
- [A Regex JIT That Compiles in 5 Microseconds](https://saksham.digital/blog/11_jit_compiling_in_5_microseconds) ([markdown](https://saksham.digital/blog/11_jit_compiling_in_5_microseconds.md)): 2026-08-24 — A JIT compiler that turns a regex into machine code in about 5 microseconds, and the generated code runs within 20% of hand-written assembly.
- [A Free Mystery Model Beat GPT-5.6 on Coding. Nobody Has Audited the Number Yet](https://saksham.digital/blog/10_ox_alpha_free_mystery_model_beat_gpt56) ([markdown](https://saksham.digital/blog/10_ox_alpha_free_mystery_model_beat_gpt56.md)): 2026-08-22 — On August 20, a model called “ox-alpha” showed up on OpenRouter with a one-week free window and zero listed pricing.
- [Cerebras CS-4's Most Important Spec Is a 2 Microsecond Interconnect](https://saksham.digital/blog/09_cerebras_cs4_the_number_nobody_is_talking_about) ([markdown](https://saksham.digital/blog/09_cerebras_cs4_the_number_nobody_is_talking_about.md)): 2026-08-20 — Cerebras announced the CS-4 this week: 43.2 petabytes per second of memory bandwidth, 250 PFLOPS of compute, three WSE-3 Turbo wafers per system.
- [GPT-5.6 Closed a 30-Year Math Gap, and the Lean Proof Is the Real Result](https://saksham.digital/blog/08_gpt56_closed_a_30_year_math_gap) ([markdown](https://saksham.digital/blog/08_gpt56_closed_a_30_year_math_gap.md)): 2026-07-21 — GPT-5.6 Sol Pro proved an Omega(d^2) lower bound on derivative-free convex optimization, closing a gap open since 1996, in one 148 minute session.
- [GPT-5.6 Beats Claude by 13 Points on One Benchmark, Loses by 16 on Another](https://saksham.digital/blog/07_gpt56_three_tiers_three_answers) ([markdown](https://saksham.digital/blog/07_gpt56_three_tiers_three_answers.md)): 2026-07-13 — Sol, the top GPT-5.6 tier, scores 53.6 on Agents’ Last Exam, 13.1 points above Claude Fable 5.
- [NVIDIA Dropped the Draft Model and Got 4x Throughput](https://saksham.digital/blog/06_nvidia_killed_the_draft_model) ([markdown](https://saksham.digital/blog/06_nvidia_killed_the_draft_model.md)): 2026-07-10 — Nemotron-Labs-Diffusion accepts 6.82 tokens per speculative step, versus 2.75 for Eagle3.
- [The 518-Token Bug: When GPT-5.5's Reasoning Gets Cut Off at Fixed Intervals](https://saksham.digital/blog/05_the_518_token_bug) ([markdown](https://saksham.digital/blog/05_the_518_token_bug.md)): 2026-07-06 — GPT-5.5’s Codex model truncates reasoning at 516, 1034, 1552 tokens: always 518 apart. One user logged 4,454 truncations in a month.
- [We Built an AI Community OS in 48 Hours. We Won.](https://saksham.digital/blog/04_communeos_we_built_an_ai_community_os_and_won) ([markdown](https://saksham.digital/blog/04_communeos_we_built_an_ai_community_os_and_won.md)): 2026-06-27 — Six AI agents. A live SQLite backend.
- [The Average Latency Lie: What 100 LLM API Calls Taught Me About Benchmarking](https://saksham.digital/blog/the-average-latency-lie) ([markdown](https://saksham.digital/blog/the-average-latency-lie.md)): 2026-06-18 — I benchmarked an LLM API 100 times. The mean said 1.7 seconds. The median said 2.2 seconds. If that ordering looks impossible to you, good.
- [I Built a Futures Market Inside a GitHub Repo](https://saksham.digital/blog/04_gitrade_futures_market_in_a_repo) ([markdown](https://saksham.digital/blog/04_gitrade_futures_market_in_a_repo.md)): 2026-06-09 — You can short my own repo. From inside the repo. Using GitHub Issues as orders.
- [Smart Pointers in C++: The Complete Guide](https://saksham.digital/blog/02_smart_pointers_complete_guide) ([markdown](https://saksham.digital/blog/02_smart_pointers_complete_guide.md)): 2026-05-20 — Every C++ developer learns about unique_ptr and shared_ptr.
- [Why I Rebuilt My HFT Matching Engine From Scratch and Got 3.3x Faster](https://saksham.digital/blog/01_why_i_rebuilt_my_hft_engine) ([markdown](https://saksham.digital/blog/01_why_i_rebuilt_my_hft_engine.md)): 2026-05-10 — I had a working order matching engine. 2 million orders. Sub-microsecond latency. It worked. But it was built wrong.

## wiki

- [Cache Locality](https://saksham.digital/blog/wiki/cache-locality) ([markdown](https://saksham.digital/blog/wiki/cache-locality.md)): 2026-10-10 — Summary Cache locality refers to how well a program’s memory access patterns exploit CPU cache hierarchy (L1/L2/L3).
- [C++ Systems Programming](https://saksham.digital/blog/wiki/cpp-systems-programming) ([markdown](https://saksham.digital/blog/wiki/cpp-systems-programming.md)): 2026-10-10 — Summary Systems programming in C++ focuses on writing code that interacts closely with hardware: manual memory management, cache-aware data structures, lock-free concurrency, and zero-overhead abstractions.
- [HFT Matching Engine: Build Log and Lessons](https://saksham.digital/blog/wiki/hft-matching-engine) ([markdown](https://saksham.digital/blog/wiki/hft-matching-engine.md)): 2026-10-10 — Summary Built a limit order book matching engine in C++ across three architecture iterations, going from 823K to 2.72M orders/sec (3.3x throughput) and 27.4μs to 3.1μs p999 latency (8.8x tail improvement).
- [Order Book Design](https://saksham.digital/blog/wiki/order-book-design) ([markdown](https://saksham.digital/blog/wiki/order-book-design.md)): 2026-10-10 — Summary A limit order book (LOB) maintains sorted lists of buy (bid) and sell (ask) orders at each price level.

## About the author

Chimera is written and maintained by Saksham Arora. Saksham Arora (nerfsaksham): data scientist, quant researcher and AI engineer from India. Builds Margin, Latch and Claude Code mods, ML pipelines and quantitative research, with a low-latency C++ background. The complete, authoritative profile — projects, skills, achievements — lives at https://saksham.digital/llms.txt and https://saksham.digital/profile.json.

- https://saksham.digital
- https://github.com/saksham10arora-dotcom
- https://linkedin.com/in/saksham-arora10
- https://x.com/nerfsaksham
- https://github.com/nerfsaksham
- https://huggingface.co/nerfsaksham
- https://www.kaggle.com/nerfsaksham
- https://dev.to/nerfsaksham
- https://medium.com/@nerfsaksham
- https://hashnode.com/@nerfsaksham
- https://www.youtube.com/@nerfsaksham
- https://bsky.app/profile/nerfsaksham.bsky.social
- https://devpost.com/nerfsaksham
- https://devfolio.co/@nerfsaksham
- https://lablab.ai/u/@nerfsaksham
- https://cerebralvalley.ai/u/nerfsaksham
- https://www.producthunt.com/@nerfsaksham
- https://pypi.org/user/nerfsaksham/
- https://www.npmjs.com/~nerfsaksham
- https://topmate.io/nerfsaksham
- https://www.instagram.com/nerfsaksham/
- https://leetcode.com/u/nerfsaksham/
- https://codeforces.com/profile/C0SBYYY
- https://www.codechef.com/users/c0sbyyy