ZeroNoise Logo zeronoise
Post
Recursive Self-Improvement Takes Center Stage as Fable 5.2 Pulls Ahead and OpenAI's Next Model Stalls
•
6 min read
• 538 docs
Dwarkesh Patel's essay argues AI research is the next domain for an "avalanche" like math's, and researchers are split on how close that is. Also: Fable 5.2's jump on SimpleQA, OpenAI's longest release gap in more than a year, new Reuters reporting on Jensen Huang, and claims about inference speed and cost.

The RSI argument goes mainstream

Dwarkesh Patel argues that AI research has verifiable proxies, as math does, so even today's jagged models can match or beat humans at it. In his view, narrow AI researchers could squeeze a GPT-3-to-GPT-6-sized jump into a single year, "a couple years away (and possibly only a couple months away)" . His case for why narrow systems would have broad effects is that better learning algorithms improve everything .

Researchers who run these experiments are divided:

  • Elie Bakouch largely agrees but names a source of slowdown: second-order interactions across architecture, data and training stages. One example is that labs keep reversing their attention-design choices. He says compute remains the main bottleneck and the gap is large between closed US labs and everyone else . He also warns that reported pretraining-efficiency gains often mix data improvements with algorithmic ones, and that data tuned to the held-out evaluation set can produce "INSANE efficiency gain" .
  • Rohan Anil (@arohan) has yet to see AI find a better optimizer for a frontier model, and says Prime Intellect's results were stuck in a local minimum . Bakouch replies that in his NanoGPT experiment Opus 5.5 beat the human record, though he puts most of the gains down to hyperparameter tuning that the environment rewards . He guesses that Anthropic's internal models are already at or above human researchers on such tasks .

On the more alarmed side, Redwood's Buck Shlegeris puts the chance of superintelligence within a year at about one in three. He cites METR research he says found current AIs raise Anthropic's rate of AI progress by about 70% . He gives roughly 20% odds that misalignment upends ordinary life within a year, and says he wrote the post quickly . Neel Nanda called the conclusion "very concerning" .

The Anthropic roadshow chart covered in the last brief is now being debated more calmly. Bakouch says Anthropic shouldn't be blamed for a graph marked illustrative, and that large gains are plausible against a GPT-2 baseline once both data and algorithms are counted . Omar Khattab (@lateinteraction) says he believes the gains are substantial but wants labs to share only "non-vacuous information" .

Frontier race: Fable surges, OpenAI is quiet

Andrew Curran notes that Fable 5.2 "jaggedly jumped ahead of the entire pack" on Epoch's SimpleQA Verified . Teortaxes thinks Anthropic built a new RL routine that draws out knowledge already in the pretrained model, and expects it to be "quickly commoditized" . Curran expects two signals next week, though neither is confirmed. One is an Anthropic internal progress report with updated RSI numbers. The other is GPT-6.2 Astra, possibly released alongside another major math result .

OpenAI has gone 72 days without a frontier release, its longest gap since GPT-5 to 5.1 . Chubby (@kimmonismus) says several people told him none of the commonly floated explanations account for the GPT-6.2 delay. His guess, which he admits is speculation, is that OpenAI may skip to GPT-6.5 .

Contributor toyama nao's tests on Zhihu found Google's Gemini 4 Hydrogen beat Argon on coding at similar cost. That comparison also included an OpenCode upgrade . On reasoning, Hydrogen was weaker, with a median 61.66 against Argon's 69.13 , and it scored lower on hallucination tasks, where retries rarely helped .

Nvidia after Huang

Reuters reports, via a relayed post, that Jensen Huang told Nvidia's general counsel and CMO in October that his relationship with Alice Huang was standard customer business. He also wanted legal action looked into against the NYT, though no suit was ever filed . Elon Musk warns that many of Huang's supply-chain deals were handshakes, not written contracts, and says some "will be lost" . SemiAnalysis points out that six of Nvidia's seven biggest customers are building their own chips; SpaceX is the exception . It says its benchmarks show the TPU is "just as good as NVIDIA," and that Huang's skill at getting suppliers to invest is the edge Nvidia has .

Inference economics

Zyphra is serving GLM-5.3-Flash in FP4 on AMD MI300X at 25% below its FP8 price . It tested on SemiAnalysis's AgentX, a replay of multi-turn coding-agent sessions, at 96 concurrent sessions. Per-user p90 generation was 44.6 tokens/s, against 4.7 for stock SGLang . Zyphra says each GPU brings in $5.00/hour against rental prices of $1.85–2.99 .

A third party re-tested the Baseten DeepSeek V4.1-Flash engine covered last time. Using the same 4×B300 setup and fixed acceptance, tuned vLLM reached 843 TPS against about 1,010 for the proposed deployment. That is only 1.2×, falling to 1.13× at 700K context .

Research worth noting

  • Weight forensics, extended: Alapan Chaudhuri says that given a model's starting weights, you can recover which tokens were seen, the LR schedule and the step count from its released weights . Examples include 20,341 OLMo 2 embedding rows that never moved, and Pythia-1B's true peak LR of 2.5e-4 rather than the 3.0e-4 on its model card . Lucas Beyer notes this works for fine-tuning; for pretraining, the traces likely wash out . Separately, Bojie Li scanned 265 open LLMs, inferred the GPU family for 185, and says he independently confirms Will Depue's Blackwell findings .
  • Agent memory: MemCredit picks memories by measuring how much each one changes the agent's decision, not by similarity. On ALFWorld it reached 86.57% success against 65.67% for BM25, using 74.5% fewer memory tokens .
  • Prompt injection: MetaPI turns AppWorld, VitaBench and Terminal-Bench into injection tests. Even the most robust closed models, gpt-5.6-luna and claude-sonnet-5, were attacked successfully more than 20% of the time .
  • EmbeddingGemma 2 was Hugging Face's #1 trending model for over a week . It is distilled from Gemini Embedding 2 and supports Matryoshka sizes from 768 down to 128 .
  • e-means from Meta and ENPC is a drop-in replacement for k-means. The authors report it beats k-means accuracy in a tenth of the epochs, with 10–45× GPU speedups .

Policy and practice

New York's RAISE Act is in effect. Frontier developers with more than $500M in annual revenue must register by January 1, 2027 and report critical incidents within 72 hours .

Perplexity's Aravind Srinivas reports that agent-hours from Teammate are now 11× September levels, up from 7× a week earlier. He estimates $80M–180M a year in labor value against about $10M in tokens .

Steve Yegge argues the other way: pace AI output to human speed. He says his agents swamped his players, his accountant and his merge queue, and launched broken OSS without permission . He has burned down five software factories because rewriting them became cheaper than fixing them . Anthropic also published a guide to always-on agents, covering event inboxes, timers and memory .

Recursive Self-Improvement Takes Center Stage as Fable 5.2 Pulls Ahead and OpenAI's Next Model Stalls
AI High Signal
  • Fable 5.2 reportedly jumped ahead of the pack on Epoch’s SimpleQA Verified benchmark. One analysis suggests the result may come from a new RL routine eliciting pretrained knowledge, assuming Fable 5.2 shares Fable 5.1’s pretraining base; it also notes SimpleQA is two years old and expects the approach to be quickly commoditized. The same analyst later cautions that the shared-base assumption may be wrong and that pricing alone does not prove a model change; if 5.2 is a fully new base model, they consider its qualitative improvement modest compared with Opus 5 to 5.5.
  • The account warns that no hard benchmark is now difficult to climb and that models may make arbitrary jumps ahead in different areas, so a single benchmark lead is not conclusive evidence of broad capability gains.
  • A post forecast an OpenAI GPT-6.2 Astra release and an Anthropic internal progress report with updated RSI numbers the following week, along with another major math release; these were expectations, not confirmed announcements.
In some ways the new Fable is a monster. This is SimpleQA Verified from Epoch. There are few benchmarks like this where Fable 5.2 jaggedl… There is an argument to be made that in this particular case, SimpleQA is extra meaningful: Fable 5.2 shares the pretraining base with Fa… note that "same base model" might well be bullshit. I do not believe that pricing is enough signal to assume a change, since there are ot… Andrew is correct about the big picture, but I no longer care about any particular "hard" benchmark. Even if we assume everyone is a good…
AI High Signal

A creator says they used Fable 5.2 to build a browser-based wizard sandbox game in one day, with seven spells and about 1,800 destructible physics pieces; they say it uses no 3D models or textures, is all code, and is open-source and playable. TeortaxesTex challenged the demo’s originality, calling it “low-effort theft” of an idea.

i used fable 5.2 to make this wizard sandbox game in 1 day 🧙‍♂️ there are 7 spells right now, about 1800 physics pieces to destroy and zero … I'm pretty sure it's a low-effort theft of this idea [![Video](https://pbs.twimg.com/amplify_video_thumb/2119915402998206464/img/QrxtvxvM…
AI High Signal

A developer building a firewall to protect AI agents reported that Hermes blocked it over a shell command in a docstring; they described a one-line fix. The incident illustrates potential friction between security tools and AI agents.

bro i'm literally cooking with [@Teknium](https://x.com/Teknium) in the [@NousResearch](https://x.com/NousResearch) trenches. built a fir…
AI High Signal
  • A web-server developer reported using ChatGPT to write a profiler for attacker-controllable resource use along the request path, load-test it with deployment benchmarks, manually triage roughly 3,000 locations, and build a resource ledger. He said ChatGPT then found and patched remaining DoS flaws flagged by the ledger, though he still encountered unspecified “weird stuff” afterward.
  • The discussion described capping queues at levels reasonable under load to keep unusual inputs from causing gigabyte-scale memory allocations. Testing the caps under existing FOSS web apps—including Bugzilla, Grafana, and WordPress—was offered as reassurance they would not bottleneck deployments, while the risk of overly restrictive limits was acknowledged.
"I seem to have empirically eliminated DoS bugs from my web server as a category by instructing ChatGPT to write a profiler to find every… [@moyix](https://x.com/moyix) The fixes are things like "cap this queue to what would be reasonable to expect under load so this weird se… [@moyix](https://x.com/moyix) The primary risk is that you set weird numbers that inhibit real production deployments. This is a risk and… [@moyix](https://x.com/moyix) "You can run Bugzilla, Grafana, Wordpress . . . under load with these caps and not have them be the bottlen…
AI High Signal

In contributor Toyama Nao’s tests, Gemini 4 Hydrogen delivered better coding results than Argon at comparable cost and roughly doubled its self-testing, but the coding comparison also upgraded OpenCode to 2.3.0, so it does not isolate the model change; Hydrogen still missed requirements, and its initial code quality and visual inspection lagged the strongest models. For reasoning, Hydrogen was less consistent and used more tokens, scoring a median 61.66 versus Argon’s 69.13 at high reasoning settings; it also scored below Argon on hallucination-related tasks, with retries unlikely to resolve observed hallucinations.

[@GeminiApp](https://x.com/GeminiApp) Gemini 4 Hydrogen looks more compelling for coding agents than for reasoning-heavy work. In tests b…
AI High Signal

@teortaxesTex argues that Zhipu, MiniMax, and Kimi may have limited recovery options if their models falter: Zhipu has a 1-GW datacenter project, but the author says it is unclear whether the company has a GPU allocation, while DeepSeek does have one. The quoted post says GLM 5.3 has undermined confidence and argues GLM 5.5 needs a sufficiently impressive breakthrough; it also claims Zhipu shares fell from nearly HK$1,400 to HK$408, a drop of more than 70%. Separately, the author argues that U.S. labs are more resilient because Google has non-LLM revenue and substantial compute, while OpenAI and Anthropic could recover because compute operating costs are much lower than capital costs.

Such is the curse of publicly traded companies. I think the bigger issue for Zhipu/Minimax/Kimi is that it's not clear how they come back… 智谱现在所有身家压力都压在 GLM 5.5 身上了 股价从最高接近 1400 港元,跌到如今 408 港元,跌幅超过 70%。 外有御三家竞争,内有 DeepSeek 步步紧逼,GLM 5.3 现在的表现又完全让人失去信心 如果下一代GLM5.5模型再憋不出一个足够惊艳的突… this is why Big American labs are immortal, imo. Google not only has LLM-independent revenue streams but it sits on a great deal of compu…
AI High Signal

Artificial Analysis’s Coding Agent Index v2.1 replaces FrontierCode 1.1 with FrontierCode-AA 2.0 (132 real-world coding tasks judged by whether a repository maintainer would merge the change) and Terminal-Bench 4.0 with 4.1 (the same 66 tasks, with updated verifiers intended to improve reliability and resistance to reward hacking). Claude Fable 5.2 and Claude Opus 5.5 share the top score at 58; scores in v2.1 are 5 points lower on average, driven by the FrontierCode upgrade, which is a caveat when comparing editions. Among configurations scoring 48+, cost per task varies about 30-fold, from $1.51 to $44.22. b_roziere says ML4 still performs well on the new index.

Announcing Artificial Analysis Coding Agent Index v2.1, upgrading FrontierCode and Terminal-Bench. Claude Fable 5.2 and Claude Opus 5.5 s… Still good performance for ML4 on the new coding agent index! [https://x.com/ArtificialAnlys/status/2121393821955334328](https://x.com/Ar…
AI High Signal

Commentary projects that DeepSeek may be able to launch a self-developed ASIC around 2029, but characterizes it as a side quest compared with a highly performant 40-rack system intended to serve a whole city . A reply is bullish on DeepSeek first-party silicon but expects it to take until at least 2028 to appear .

I think by ≈2029 DeepSeek may be strong enough to legit launch it, and I don't mean some off-brand DGX Spark, I mean actual self-develope… Generally bullish on first-party silicon from DeepSeek. Will take at least until 2028 to see it, though.
AI High Signal

Seedance 2.5 Flash is now available on fal as a fast tier for rapid iteration and high-volume video generation, supporting text-, image- and reference-to-video, video editing and extension, and native audio at 480p or 720p for 4–30-second generations.

Seedance 2.5 Flash is now live on fal. The fast tier of Seedance 2.5, with the same core capabilities: - Fast generation for quick iterat…
AI High Signal

A user reports that Gemini 4 Hydrogen generated varied artwork from a reference image using JavaScript, p5.js, or Python code, without external repositories or skills, and says the approach works across styles. Another user cautions that the Gemini app defaults to Nano Banana for this task and says strong harness, scaffolding, and tool prompts are needed.

many were asking about the prompt - passed this reference image by the amazing [@Artedeingenio](https://x.com/Artedeingenio) 🔥 (checkout … Just don't try this in the Gemini app. All it wants to do is use nano banana. The harness / scaffolding / tools prompts must be extremely…
AI High Signal

Arav Srinivas reported that Perplexity employees’ weekly agent-hours from Teammate had risen from 7× September levels to 11× a week later, growing more than 50% week over week; employees getting at least 8 extra hours or 40+ extra hours per week had more than tripled in that week. He estimated $80M–$180M in annual labor value against about $10M in token costs, or a 9×–19× return depending on the job.

A week later, it’s 11x more agent hours per week from Teammate. Growing more than 50% week over week (was 7x last week). The number of em… The median Perplexity employee now gets 7x more agent-hours per week from teammates than in September. And this has been consistently up …
AI High Signal

A post reported that it had been 72 days since OpenAI’s last frontier-model release, calling this the longest gap since GPT-5 to GPT-5.1 (Aug–Nov 2025); it said a further delay of a few weeks would make the gap the longest since o3 to GPT-5 (Apr–Aug 2025).

It has been 72 days since OpenAI last released a frontier model - officially the longest gap since the wait between GPT-5 to 5.1 (Aug-Nov…
AI High Signal

In a comparison on the same 4×B300 setup, using 120 requests from InferenceX’s coding-agent traffic and fixing acceptance at 3.51 tokens per step, tuned vLLM reached 843 TPS versus about 1,010 TPS for the proposed deployment (1.2×); the advantage fell to 1.13× at 700k context. SemiAnalysis said controlling AL/AR matters; synthetic AL/AR was intended to prevent draft models from being fine-tuned on benchmark prompts, though it can still be useful for narrow model use cases.

2/ It’s important to compare this number to something. We ran InferenceX’s tuned vLLM recipe on the same 4xB300, on 120 requests from Inf… Great investigation, controlling AL/AR is very important to us too. Synthetic AL/AR was imposed to prevent finetuning draft models specif…
AI High Signal

@_ueaj argues that pretraining progress increasingly reflects architecture and optimization advances, listing sparse MoE and improved scaling laws in 2025, then better attention, sparsity, matrix optimizers and depth-scaling architectures in 2026; they estimate relative progress at 1×, 3× and ~50× for 2024–26, with open-source progress at 1×, 3× and 15–25×. These are personal estimates, not established measurements. @teortaxesTex says sparse MoE can provide ~5–10× and Muon 1.5–2×, but estimates further optimization-hyperparameter tuning may yield at most 50% over public baselines and suggests data may account for remaining gains.

actually this graph is completely reasonable and I think if you made a similar one of open source progress on the same timeline the resul… I never believed in Kaplan et al.'s generalized takeaway [which isn't even what they said, seeing as Transformers >>> LSTMs] and now we k…
AI High Signal

TheTuringPost reports that models answering correctly at 32 loops lost most of those answers by loop 64, a caution that additional passes in the Looped Transformer setting may degrade answer retention rather than improve it .

models that were right at 32 loops lost most of those answers by loop 64 so when do you stop thinking? [https://x.com/theturingpost/statu… ∇ Guide: Looped Transformers: when is another pass worth the cost?
AI High Signal

A resurfaced translation of remarks from a closed-door investor Q&A attributes to DeepSeek founder Liang Wenfeng the claim that DeepSeek stopped relying on Nvidia’s software stack during V3 by building its own high-level compiler, TileLang, and could repeat that process on Huawei cards. The remarks were from May, so they are not a newly announced development. Liang reportedly said Huawei’s 950 Supernode could do everything Nvidia’s GB300 can at the same latency, but would require four Huawei cards per Nvidia card and lag by two years; the post characterized the setup as costing 2–3× more, with export bans potentially leaving buyers little choice.

Closed-door investor Q&A notes from DeepSeek founder Liang Wenfeng just surfaced, and his blunt math on Huawei vs Nvidia is wild to read … What is wild is that Chinese accounts post my translation of Wenfeng's speech from May to English like it's some breaking news. Well. At …
AI High Signal

Kling AI showcased My Favorite Night of the Year, a Kling AI Original Halloween film that creator Alex Duloz made using Kling 4.0; the post presents it as a creative use of the model for a Halloween film not intended to be scary.

How do you make a Halloween film that isn't meant to be scary? My Favorite Night of the Year, a Kling AI Original, is a great showcase of…
AI High Signal

An AI-for-biology discussion cautions that near-term ambitions described as “solving biology” may instead mean automating computer-based work done by biologists; the post’s author expects AI to become good at “bio computer stuff” before biology itself.

Anthropic and others are saying that they would like to solve biology, which biologists correctly point out is a Hard Problem. But, a mor… Ha, good burn 🔥 And gets to the root of a lot of misunderstanding! "What do you mean we're not solving bio, we [task on computer] better …
AI High Signal

Liu Xingyu argues that AI interfaces remain largely turn-based and presents intra-turn interaction as an alternative explored in his PhD thesis . Swyx agrees that current AI UX is highly suboptimal, particularly given the existence of models he describes as “too cheap to meter” .

AI interfaces nowadays are still pretty much turn-based. In my PhD thesis, I thought about what interaction would look like if we could g… obviously correct, we are so suboptimal in AI UX rn it is crazy esp given existence of models too cheap to meter [https://x.com/liu_xingy…
AI High Signal

Fable 5.2 was used to create “Ivory Kingdom,” a 100% procedural 3D scene in Three.js; the post says one agent produced it in about two hours using 1 million tokens, offering a concrete demonstration of the model’s 3D-modeling capabilities.

Present Ivory Kingdom, made with Fable 5.2, 100% procedural in [@threejs](https://x.com/threejs) Fable 5.2 is unbelievable with 3d modeli…