# Poetiq

> Source: https://aiwiki.ai/wiki/poetiq
> Updated: 2026-10-27
> Fact-checked: 2026-10-27
> Categories: AI Agents, AI Companies, AI Research, Large Language Models
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "Poetiq." aiwiki.ai, 27 Oct 2026. https://aiwiki.ai/wiki/poetiq
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

**Poetiq** is an American artificial intelligence startup, based in Mountain View, California, that builds software on top of existing [large language models](https://aiwiki.ai/wiki/large_language_model) instead of training its own. It was founded in June 2025 by Shumeet Baluja and Ian Fischer, two former researchers at [Google DeepMind](https://aiwiki.ai/wiki/google_deepmind) who run the company as co-CEOs.[1][2] Poetiq's core product is what it calls a "meta-system" (styled "Metasystem" in its later posts): software that automatically writes and tunes a task-specific [harness](https://aiwiki.ai/wiki/harness) of code, prompts and search strategies around models such as Gemini, GPT and Claude. The company describes this as a form of [recursive self-improvement](https://aiwiki.ai/wiki/recursive_self-improvement) and says its goal is [superintelligence](https://aiwiki.ai/wiki/superintelligence).[1][18]

The company became known in late 2025 for its results on the [ARC-AGI](https://aiwiki.ai/wiki/arc_agi) reasoning benchmarks. In December 2025 the ARC Prize team verified an open-source Poetiq harness built on [Gemini 3 Pro](https://aiwiki.ai/wiki/gemini_3_pro) at 54% on the semi-private evaluation set of [ARC-AGI-2](https://aiwiki.ai/wiki/arc_agi_2), at a cost of $30.57 per task, which placed it at the top of the ARC Prize leaderboard at the time.[6][7][12] In January 2026 Poetiq announced a $45.8 million seed round co-led by FYRFLY Venture Partners and Surface Ventures.[2][3]

## Overview

| Field | Detail |
| --- | --- |
| Type | Private company (Poetiq Inc.)[3] |
| Founded | June 2025[2] |
| Founders | Shumeet Baluja, Ian Fischer (co-CEOs)[1][2] |
| Headquarters | Mountain View, California, United States[2] |
| Focus | Automatically built harnesses that improve the reasoning, accuracy and cost of existing LLMs[2][17] |
| Main product | Poetiq meta-system ("Metasystem"); Augur model-mapping tool[17][19] |
| Open-source release | ARC-AGI solver (MIT license)[13] |
| Funding | $45.8 million seed (January 2026)[2] |
| Valuation | Not disclosed[4] |
| Website | poetiq.ai |

## History

### Founding

Baluja and Fischer had worked together for more than a decade at Google before founding Poetiq, according to Puck.[4] In the company's funding announcement, Poetiq said the two had noticed while collaborating at Google DeepMind that frontier LLMs "were struggling to solve most hard (or easy) problems," and that the usual remedy, pre-training and post-training models through [reinforcement learning](https://aiwiki.ai/wiki/reinforcement_learning), takes weeks and is too expensive for all but the largest companies.[2] The company was founded in June 2025.[2] In an August 2026 essay Poetiq wrote that it had started "just over a year ago," with its system first optimizing pieces of individual tasks.[18]

Puck reported in January 2026 that, unlike most early-stage AI companies, Poetiq did not train or fine-tune LLMs, develop specialized models or buy GPUs, and that it had reached its ARC-AGI-2 result "in just six months, with a six-person team, and a total hardware bill of only $40,000."[4]

### ARC-AGI results (November and December 2025)

On November 20, 2025, Poetiq published its first public results in a post titled "Traversing the Frontier of Superintelligence." It claimed new cost-versus-accuracy frontiers on the public evaluation sets of both ARC-AGI-1 and ARC-AGI-2.[5] The headline configurations used [Gemini 3](https://aiwiki.ai/wiki/gemini_3) and [GPT-5.1](https://aiwiki.ai/wiki/gpt-5.1), released on November 18 and November 13, 2025, respectively. Poetiq also showed cheaper systems built on xAI's Grok 4 Fast and on the open-weight [gpt-oss](https://aiwiki.ai/wiki/gpt_oss)-120B model, the latter for less than one cent per problem on ARC-AGI-1.[5] The company said all of its meta-system's adaptation had been done before Gemini 3 and GPT-5.1 were released, using only open-source models, and that the system had never been shown ARC-AGI-2 problems.[5]

The same post said Poetiq's ARC-AGI-2 public-set results exceeded the 60% scored by the average human test-taker, while noting that most underlying models had scored lower on the held-out semi-private set of ARC-AGI-1 than on its public set, and that it expected a similar drop.[5] The Decoder made the same point in its coverage: the scores applied only to public data, which can end up in the training data of large models.[11] Poetiq released the code for its Gemini 3 configurations on GitHub under the MIT license.[5][13]

On December 5, 2025, Poetiq reported that the ARC Prize team had evaluated its open-source solver on the ARC-AGI-2 semi-private test set and measured 54% at $30.57 per problem. The previous best, Gemini 3 Deep Think, had scored 45% at $77.16 per problem.[6] ARC Prize's 2025 results post, published the same day, said it had created a new "Model Refinements" leaderboard category and verified the Poetiq harness, which raised Gemini 3 Pro from a baseline of 31% at $0.81 per task to 54% at $31 per task. ARC Prize added that, as reported by Poetiq, the same refinement gave similar gains on [Claude Opus 4.5](https://aiwiki.ai/wiki/claude_opus_4_5) at about twice the cost, roughly $60 per task.[7] ARC Prize announced the verification on X on December 7, 2025.[9]

On December 23, 2025, Poetiq reported that the same harness, run with the X-High variant of OpenAI's [GPT-5.2](https://aiwiki.ai/wiki/gpt-5.2), reached "as high as 75%" on the full ARC-AGI-2 public evaluation set at under $8 per problem, about 15 percentage points above the previous state of the art.[10] This figure is from the public set; the ARC Prize leaderboard lists only the Gemini 3 Pro configuration as a verified Poetiq entry.[39] According to Poetiq's funding announcement, OpenAI president [Greg Brockman](https://aiwiki.ai/wiki/greg_brockman) responded that Poetiq was "exceeding the human baseline on ARC-AGI-2 with gpt-5.2."[2]

### Seed funding (January 2026)

On January 29, 2026, Poetiq announced that it had raised $45.8 million in seed funding. FYRFLY Venture Partners and Surface Ventures co-led the round, with participation from [Y Combinator](https://aiwiki.ai/wiki/y_combinator), 468 Capital, Operator Collective, Hico Ventures and Neuron Venture Partners.[2][3] Puck, which reported the round the same day, said it was raised "at an undisclosed valuation."[4] The press release described Poetiq as emerging from stealth; the company's own later posts date its emergence from stealth to November 2025.[2][15] Outset Capital, which is not named in the press release, also describes Poetiq as a portfolio company that it backed "at the earliest stage."[22]

"That Poetiq managed to top ARC-AGI within six months of launching is remarkable," Philipp Stauffer, a general partner at FYRFLY, said in the announcement.[2] Poetiq said it would use the money to run its "recursive improvement loops around the clock" and to grow its team.[14]

### Benchmark program (2026)

After ARC-AGI, Poetiq said it picked benchmarks to cover three kinds of task: reasoning (ARC-AGI), retrieval of knowledge stored in a model's weights ([Humanity's Last Exam](https://aiwiki.ai/wiki/humanity_s_last_exam) and [SimpleQA Verified](https://aiwiki.ai/wiki/simpleqa_verified)), and coding (LiveCodeBench Pro).[16][17]

- **February 10, 2026: Humanity's Last Exam and SimpleQA Verified.** Combining Gemini, GPT and Claude models in its harness, Poetiq reported 55.0% on Humanity's Last Exam with tools, against 53.1% for [Claude Opus 4.6](https://aiwiki.ai/wiki/claude_opus_4_6). Without tools, it reported 42.5% using Gemini 3 Pro, against Claude Opus 4.6's 40.0%. On SimpleQA Verified it reported 77.3% for an ensemble of models, against 72.1% for Gemini 3 Pro alone.[15] Poetiq said it integrated Claude Opus 4.6 into the harness within 24 hours of the model's February 5 release.[15]
- **February 27, 2026: Lightcone interview.** Fischer discussed the company's "reasoning harnesses" on Y Combinator's Lightcone video series, in an episode titled "The Powerful Alternative To Fine-Tuning."[24]
- **May 14, 2026: LiveCodeBench Pro.** Poetiq said its meta-system built a coding harness from scratch for the 25Q2 edition of LiveCodeBench Pro, optimizing it only with [Gemini 3.1 Pro](https://aiwiki.ai/wiki/gemini_3_1_pro). The harness raised Gemini 3.1 Pro from 78.6% to 90.9%. Applied without further optimization to [GPT-5.5](https://aiwiki.ai/wiki/gpt-5.5), it reached 93.9%, up from 89.6%.[16] The company stressed that this involved no [fine-tuning](https://aiwiki.ai/wiki/fine_tuning) and no access to model internals, only standard API calls.[16]
- **July 15, 2026: "Benchmarks Are Dead (for us)."** Poetiq said its meta-system, with "zero human intervention," built harnesses that set state-of-the-art results on six benchmarks it had not previously seen, covering competition mathematics, scientific coding, long-horizon planning, agentic tool use and long-context retrieval.[17]
- **August 2026: research essay and Augur.** On August 4 the company published "A Poetiq Perspective on Recursive Self-Improvement," setting out its technical roadmap and safety arguments, and on August 10 it introduced Augur, a tool for mapping where models fail.[18][19]
- **October 15, 2026: LLM router.** Poetiq reported that its meta-system had built and optimized LLM routers for three simulated enterprise workloads.[20]

## Technology

### The meta-system

Poetiq's homepage summarizes its approach as "Poetiq builds on top, not into the model." The company describes LLMs as large stores of knowledge whose contents are only partly reachable through naive prompting, and says the work is to figure out "what questions to ask, not just how to ask them" and to piece the resulting fragments together.[1] Its pitch to customers is that they supply a problem and a few hundred examples, rather than the thousands or millions that fine-tuning or RL [post-training](https://aiwiki.ai/wiki/post-training) require. The meta-system then generates an agent specialized for that problem and improves it to be more accurate and cheaper to run.[2][3]

In its ARC-AGI write-up, Poetiq described two principles behind its open-sourced configurations. First, "the prompt is an interface, not the intelligence": the system asks the LLM for a candidate solution, sometimes as code, receives feedback, analyzes it and asks the LLM to refine the answer. Second, the system audits its own progress and decides when it has a satisfactory answer, stopping early to save cost.[5] SiliconANGLE described the same loop in its coverage of the funding round.[3] The Decoder explained the ARC-AGI solver as guiding the underlying model to write code for each task, running the code, checking it for correctness and fixing errors, with multiple independent runs combined into the final answer.[11]

The company stresses that its harnesses are not tied to one model. It says it integrated Gemini 3 and GPT-5.1 within hours of their release, and that its LiveCodeBench Pro harness, optimized for Gemini 3.1 Pro, also improved every other model it was tested on, including the open-weight [Kimi K2.6](https://aiwiki.ai/wiki/kimi_k2_6), which by Poetiq's figures went from 50.0% to 79.9%.[5][16] By July 2026 Poetiq said the meta-system designed the code, prompts and hyperparameters of each new harness itself, and that the company's role had shrunk to picking the benchmark.[17]

In its August 2026 essay, Poetiq called the meta-system a "self-optimizing optimizer": each task it optimizes is also used to improve the meta-system's own code. The company said all of its state-of-the-art results so far had come without modifying any LLM's parameters.[18] The essay compares Poetiq with other approaches described as self-improving, including Sakana AI's Darwin Gödel Machine, the SICA coding agent, Google DeepMind's [AlphaEvolve](https://aiwiki.ai/wiki/alphaevolve), and prompt-optimization frameworks such as [DSPy](https://aiwiki.ai/wiki/dspy) and GEPA. It classifies Poetiq, like the Darwin Gödel Machine and SICA, as a "self-rewriting agent" that improves the scaffolding around a frozen model, and acknowledges that such systems may be limited by the quality of the underlying LLMs.[18]

### ARC-AGI solver

The open-source repository, `poetiq-ai/poetiq-arc-agi-solver`, is a Python project released under the MIT license. Its README says it reproduces Poetiq's ARC-AGI-1 and ARC-AGI-2 submissions and can recreate the Gemini 3 points from the company's charts, given API keys for the models being tested.[13] As of October 2026 the repository had more than 1,200 stars on GitHub.[13]

### Roadmap and safety claims

Poetiq's August 2026 essay describes a three-stage technical roadmap:[18]

| Stage | Description (company's account) | Status as stated in August 2026 |
| --- | --- | --- |
| 1. Building on top of models | A recursive self-improvement loop that uses LLMs as tools and improves harnesses and its own code | "Already mature" |
| 2. Customizing open-weights models | [Post-training](https://aiwiki.ai/wiki/post-training) an RSI-focused model on data the current loop generates, such as capability maps and RL-ready training data | Next planned step |
| 3. Training RSI models from scratch | Training the company's own models from scratch | Long-term plan |

The same essay sets out three safety arguments. Poetiq says humans choose the problems the system works on and it cannot set its own goals; that customers receive narrow, task-specific harnesses while the meta-system itself is never exposed; and that its improvements can be inspected, because a harness consists of code, prompts and human-readable data rather than changes to opaque model weights.[18] These are the company's own claims.

In an August 2026 opinion piece for CIO, Fischer and Baluja argued that enterprises should not tie their AI strategy to one model provider and should instead build model-agnostic systems that improve whenever any new model is released.[28]

### Augur and LLM routing

Augur, introduced on August 10, 2026, is a Poetiq tool that generates questions to find where a model's knowledge ends and where its reasoning fails. It sorts each failure into one of three kinds: a missing fact, a fact that is contested across sources, or a reasoning error in which the model knows every underlying fact but combines them wrongly. Poetiq says knowledge failures can be turned into supervised fine-tuning data and reasoning failures into reinforcement-learning tasks, while contested items are discarded.[19] Because the questions are generated dynamically, the company presents Augur as a benchmark that cannot be memorized. In a case study of [Kimi K3](https://aiwiki.ai/wiki/kimi_k3), Poetiq reported that when the model gave a divergent answer to a multi-fact question, it knew all of the underlying facts more than 75% of the time.[19]

On October 15, 2026, Poetiq reported using its meta-system to build [LLM routers](https://aiwiki.ai/wiki/llm_model_routing), which send each query to the cheapest model likely to answer it correctly. It tested three simulated companies with different model mixes and question distributions. In the multi-provider scenario, Poetiq said its most accurate router was about 20% cheaper than always using [Claude Opus 5](https://aiwiki.ai/wiki/claude_opus_5), with no statistically significant loss of accuracy, and that its routers beat [OpenRouter](https://aiwiki.ai/wiki/openrouter)'s Auto Router and the [Jev](https://aiwiki.ai/wiki/jev) classification model on its tests.[20] The company said the project took one engineer 2.5 weeks, of which 12 hours was the meta-system's own optimization.[20] The workloads were simulated, and the results are company-reported.

## Benchmark results

The table below lists Poetiq's main published results. Only the December 2025 ARC-AGI-2 semi-private score was verified by a benchmark organizer; the others are Poetiq's own reports, with comparison figures as Poetiq stated them.

| Date | Benchmark | Poetiq result (underlying models) | Comparison cited | Verification |
| --- | --- | --- | --- | --- |
| Dec 2025 | ARC-AGI-2, semi-private set | 54% at $30.57 per task (Gemini 3 Pro) | Gemini 3 Deep Think 45% at $77.16 | Verified by ARC Prize[6][7] |
| Dec 2025 | ARC-AGI-2, public evaluation set | Up to 75% at under $8 per task ([GPT-5.2](https://aiwiki.ai/wiki/gpt-5.2) X-High) | About 15 points above previous best | Company-reported[10] |
| Feb 2026 | [Humanity's Last Exam](https://aiwiki.ai/wiki/humanity_s_last_exam), with tools | 55.0% (Gemini, GPT and Claude) | [Claude Opus 4.6](https://aiwiki.ai/wiki/claude_opus_4_6) 53.1% | Company-reported[15] |
| Feb 2026 | Humanity's Last Exam, no tools | 42.5% (Gemini 3 Pro) | Claude Opus 4.6 40.0% | Company-reported[15] |
| Feb 2026 | [SimpleQA Verified](https://aiwiki.ai/wiki/simpleqa_verified) | 77.3% (ensemble) | Gemini 3 Pro 72.1% | Company-reported[15] |
| May 2026 | LiveCodeBench Pro (25Q2) | 93.9% ([GPT-5.5](https://aiwiki.ai/wiki/gpt-5.5)); 90.9% ([Gemini 3.1 Pro](https://aiwiki.ai/wiki/gemini_3_1_pro)) | GPT-5.5 89.6%; Gemini 3 Deep Think 88.8% | Company-reported[16] |
| Jul 2026 | ArXivMath | 89.2% (GPT-5.5) | [Claude Fable 5](https://aiwiki.ai/wiki/claude_fable_5) 87.5% (released after Poetiq's run) | Company-reported[17] |
| Jul 2026 | [SciCode](https://aiwiki.ai/wiki/scicode) | 61.5% (Gemini 3.1 Pro) | Claude Fable 5 60.2% | Company-reported[17] |
| Jul 2026 | Haladir, full set (score 0 to 1) | 0.69 (Gemini 3.1 Pro) | [Claude Opus 4.8](https://aiwiki.ai/wiki/claude_opus_4_8) 0.33 | Company-reported[17] |
| Jul 2026 | Haladir, challenge set (score 0 to 1) | 0.47 (Gemini 3.1 Pro) | Claude Fable 5 0.28 | Company-reported[17] |
| Jul 2026 | MCP-Atlas | 89.8% ([Muse Spark](https://aiwiki.ai/wiki/muse_spark) 1.1 and [Gemini 3.5 Flash](https://aiwiki.ai/wiki/gemini_3_5_flash)) | Muse Spark 1.1 88.1% | Company-reported[17] |
| Jul 2026 | [MRCR](https://aiwiki.ai/wiki/mrcr) v2 | 99.26% (Gemini 3.1 Flash-Lite) | GPT-5.4 97.3% | Company-reported[17] |
| Jul 2026 | Toolathlon | 59.26% (Gemini 3.5 Flash) | Gemini 3.5 Flash 56.5% on the unmodified set | Company-reported[17] |

For Toolathlon, Poetiq noted that Anthropic had published higher scores for Claude Fable 5 (61.7%) and Claude Opus 4.8 (59.9%) on a modified version of the benchmark, with Anthropic's own caveat that those numbers run about three points above a strictly upstream-equivalent harness.[17]

## Founders and team

### Shumeet Baluja

Baluja received a bachelor's degree from the University of Virginia in 1991 and a PhD in computer science from Carnegie Mellon University in 1996.[25] According to a CIO contributor biography, he was part of the Carnegie Mellon Robotics Institute team behind the Navlab self-driving vehicles in the early 1990s.[26] Before Google he was chief technology officer of JAMDAT Mobile, chief scientist at Lycos, and senior vice president of research and development at eCompanies.[25] Poetiq's funding announcement says JAMDAT went public in 2004.[2]

Baluja then spent more than two decades at Google and DeepMind.[26] His published research spans computer vision and facial image processing, advertisement display and optimization, automated vehicle control, statistical machine learning and high-dimensional optimization.[25] Poetiq says that during his time at Google he founded its mobile practice, started its fundamental computer vision research group, contributed to more than 170 patents and was "one of the originators of [YouTube](https://aiwiki.ai/wiki/youtube)'s copyright system."[2] He is also the author of *The Silicon Jungle: A Novel of Deception, Power, and Internet Intrigue*, a novel about data mining at a fictional internet company, published by Princeton University Press in 2011.[37]

### Ian Fischer

Fischer, a former software engineer at Webroot, co-founded Apportable with Collin Jackson in 2011 and served as its chief technology officer.[29][2] Apportable, which began at Carnegie Mellon University's Silicon Valley campus, built tools that let iOS apps run on Android.[30][29] It raised $7.4 million from investors including GV and Y Combinator.[29] Google acquired the Apportable team in a deal that VentureBeat reported was believed to have closed in March 2015; Fischer officially joined Google in May 2015.[29] Operator Collective, one of Poetiq's investors, says he spent more than ten years at Google and DeepMind.[23]

At Google and Google DeepMind, Fischer's research included work on the information bottleneck.[27] He co-authored "Deep Variational Information Bottleneck" (2016), which used neural networks to approximate the information bottleneck method of Naftali Tishby and colleagues.[33] He was the sole author of "The Conditional Entropy Bottleneck" (2020), which argued that failures of robust generalization, such as vulnerability to adversarial examples, come from models retaining too much information about their training data.[34] He was also a co-author of PlaNet ("Learning Latent Dynamics for Planning from Pixels," 2018), a model-based agent that learns environment dynamics from images and plans in a latent space, and of ReadAgent (2024), an LLM agent that compresses long documents into short "gist memories" and, in the authors' experiments, increased effective context length up to 20 times.[35][36]

### Earlier joint work

Baluja and Fischer published together before founding Poetiq. In 2017 they co-authored "Adversarial Transformation Networks," which trained feed-forward networks to generate adversarial examples against target image classifiers.[31] In January 2025 both were co-authors, with Kuang-Huei Lee and others, of "Evolving Deeper LLM Thinking," which introduced Mind Evolution. That method uses an LLM to generate, recombine and refine candidate answers whenever a solution evaluator is available, a form of [test-time compute](https://aiwiki.ai/wiki/test_time_compute) scaling. The paper reported that Mind Evolution solved more than 98% of TravelPlanner and Natural Plan problem instances using Gemini 1.5 Pro without a formal solver.[32]

### Team

As of October 2026, Poetiq's website listed Yair Alon and Saurabh Singh as founding research scientists, Michael Hale and Ashwin Baluja as founding research engineers, and Michele Covell as a research scientist, alongside the two co-CEOs.[1] The company described itself as six researchers and engineers with a combined 53 years at Google DeepMind in November 2025, and as seven people with 72 combined years at Google and DeepMind in February 2026.[5][15] By July 2026 it described a "rapidly growing technical team" with backgrounds at Google DeepMind, Apple, Microsoft, Amazon and ByteDance.[17] Poetiq's posts mention work with early customers and partners but do not name them.[15][16][17]

## Funding

| Date announced | Round | Amount | Lead investors | Other participants | Valuation |
| --- | --- | --- | --- | --- | --- |
| January 29, 2026 | Seed | $45.8 million | FYRFLY Venture Partners, Surface Ventures | Y Combinator, 468 Capital, Operator Collective, Hico Ventures, Neuron Venture Partners | Not disclosed[4] |

Sources: Poetiq press release and SiliconANGLE.[2][3]

## Reception and analysis

The ARC Prize Foundation called 2025 the "Year of the Refinement Loop" and discussed Poetiq's harness in its section on refinements added at the application layer. Its results post said an important finding with frontier commercial models released in late 2025, such as Gemini 3 and Claude Opus 4.5, was that refinement loops added at the application layer could "meaningfully improve task reliability" instead of relying only on the providers' own reasoning systems, provided the base model has enough knowledge of the task domain.[7] The ARC Prize 2025 technical report repeated this and noted that the refinement harnesses it had seen were domain-specific.[8] A March 2026 academic survey of ARC-AGI approaches also cites Poetiq's 54% as an example of commercial systems scoring higher than the winning entries in the ARC Prize 2025 Kaggle competition.[38]

Coverage also questioned how far the results transfer. The Decoder reported that when an observer argued the method was specific to ARC-AGI, Poetiq replied that the ARC-AGI solver was specialized but the meta-system behind it was designed for general use.[11] Writing about the LiveCodeBench Pro result, Startup Fortune noted that Poetiq's headline numbers were company-reported and that buyers would want independent verification. It also pointed to a business risk: frontier model providers are building better agents and evaluators themselves, and if orchestration "is just a feature," Poetiq's window could narrow.[21] Ben Dickson of TechTalks read the ARC-AGI-2 result as a sign that progress in AI reasoning was shifting from scaling models toward well-engineered systems at the application layer.[12]

The verified ARC-AGI-2 record was overtaken within months as new frontier models arrived. ARC Prize's leaderboard, for example, lists Gemini 3.1 Pro (preview) at 77.1% on the semi-private set at $0.96 per task.[39]

As of October 2026, Debarghya Das's "Ultimate Neolab List," a compilation of AI research startups, included Poetiq but marked it as "Borderline," describing it as a "research team wrapping other companies' models rather than training big ones; low compute."[40]

## See also

- [List of neolabs](https://aiwiki.ai/wiki/list_of_neolabs)
- [Neolab](https://aiwiki.ai/wiki/neolab)
- [ARC-AGI-2](https://aiwiki.ai/wiki/arc_agi_2)
- [Recursive self-improvement](https://aiwiki.ai/wiki/recursive_self-improvement)
- [Harness (AI)](https://aiwiki.ai/wiki/harness)
- [Test-time compute](https://aiwiki.ai/wiki/test_time_compute)
- [LLM model routing](https://aiwiki.ai/wiki/llm_model_routing)

## References

1. Poetiq, [Poetiq homepage](https://poetiq.ai/), accessed October 27, 2026.
2. Poetiq (PR Newswire), [Poetiq Raises $45.8M for AI Meta-System, Surpasses Top LLMs on Industry Benchmark](https://www.prnewswire.com/news-releases/poetiq-raises-45-8m-for-ai-meta-system-surpasses-top-llms-on-industry-benchmark-302674571.html), January 29, 2026.
3. Maria Deutscher, SiliconANGLE, [Poetiq nabs $45.8M in seed funding for its LLM-enhancing 'meta-system'](https://siliconangle.com/2026/01/30/poetiq-nabs-45-8m-seed-funding-llm-enhancing-meta-system/), January 30, 2026.
4. Ian Krietzberg, Puck, [Poetiq Justice](https://puck.news/how-poetiqs-six-person-team-beat-google-at-ai/), January 29, 2026.
5. Poetiq, [Traversing the Frontier of Superintelligence](https://poetiq.ai/posts/arcagi_announcement/), November 20, 2025.
6. Poetiq, [Poetiq Shatters ARC-AGI-2 State of the Art at Half the Cost](https://poetiq.ai/posts/arcagi_verified/), December 5, 2025.
7. Mike Knoop, ARC Prize, [ARC Prize 2025 Results & Analysis](https://arcprize.org/blog/arc-prize-2025-results-analysis), December 5, 2025.
8. François Chollet, Mike Knoop, Gregory Kamradt and Bryan Landers, [ARC Prize 2025: Technical Report](https://arxiv.org/abs/2601.10904), arXiv:2601.10904, January 2026.
9. ARC Prize (@arcprize), [post on X announcing verification of Poetiq's Gemini 3 Pro refinement](https://x.com/arcprize/status/1997743855203148038), December 7, 2025.
10. Poetiq, [Update: Poetiq+GPT-5.2 = SOTA](https://poetiq.ai/posts/update_arcagi_gpt52/), December 23, 2025.
11. Maximilian Schreiner, The Decoder, [The ARC benchmark's fall marks another casualty of relentless AI optimization](https://the-decoder.com/the-arc-benchmarks-fall-marks-another-casualty-of-relentless-ai-optimization/), November 29, 2025, updated December 25, 2025.
12. Ben Dickson, TechTalks, [Beyond raw intelligence: How Poetiq cracked the ARC-AGI-2 benchmark](https://bdtechtalks.com/2025/12/09/poetiq-arc-agi-2-solution/), December 9, 2025.
13. Poetiq, [poetiq-arc-agi-solver](https://github.com/poetiq-ai/poetiq-arc-agi-solver), GitHub repository, accessed October 27, 2026.
14. Poetiq, [Poetiq Secures $45.8M Seed Round](https://poetiq.ai/posts/seed_funding/), January 29, 2026.
15. Poetiq, [Raising the Bar on Humanity's Last Exam and SimpleQA](https://poetiq.ai/posts/raising_the_bar_hle_simpleqa/), February 10, 2026.
16. Poetiq, [Recursive Self-Improvement Delivers New State-of-the-Art Coding Performance](https://poetiq.ai/posts/recursive_self_improvement_coding/), May 14, 2026.
17. Poetiq, [Benchmarks Are Dead (for us)](https://poetiq.ai/posts/benchmarks_are_dead/), July 15, 2026.
18. Poetiq, [A Poetiq Perspective on Recursive Self-Improvement](https://poetiq.ai/posts/rsi_perspective/), August 4, 2026.
19. Poetiq, [Augur: Model Routing, Training Data, and Benchmark Creation for LLMs](https://poetiq.ai/posts/augur/), August 10, 2026.
20. Poetiq, [Our Metasystem Built and Optimized an LLM Router](https://poetiq.ai/posts/llm_router/), October 15, 2026.
21. Startup Fortune, [Poetiq shows model orchestration can beat bigger coding models](https://startupfortune.com/poetiq-shows-model-orchestration-can-beat-bigger-coding-models/), May 15, 2026.
22. Outset Capital, [Poetiq Raises $45.8M Seed](https://www.outsetcapital.com/blog/poetiq-seed), January 2026.
23. Haley Brannan, Operator Collective, [Poetiq: The Meta-System Making AI Actually Reason](https://www.operatorcollective.com/blog-posts/poetiq-the-meta-system-making-ai-actually-reason), January 29, 2026.
24. Y Combinator, [The Powerful Alternative To Fine-Tuning](https://www.youtube.com/watch?v=UPGB-hsAoVY) (Lightcone, with Ian Fischer), YouTube, February 27, 2026.
25. Google Research, [Shumeet Baluja profile](https://web.archive.org/web/20241114043526/https://research.google/people/author35/), archived November 14, 2024.
26. CIO, [Shumeet Baluja contributor profile](https://www.cio.com/profile/shumeet-baluja/), accessed October 27, 2026.
27. CIO, [Ian Fischer contributor profile](https://www.cio.com/profile/ian-fischer/), accessed October 27, 2026.
28. Ian Fischer and Shumeet Baluja, CIO, [The enterprise AI strategy that outlasts any single model](https://www.cio.com/article/4204569/the-enterprise-ai-strategy-that-outlasts-any-single-model.html), August 4, 2026.
29. Ken Yeung, VentureBeat, [Google secretly acqui-hired mobile app development startup Apportable](https://web.archive.org/web/20161227120418/http://venturebeat.com/2016/08/18/google-secretly-acqui-hired-mobile-app-development-startup-apportable/), August 18, 2016 (archived).
30. Carnegie Mellon University, [Apportable Entrepreneurship](https://www.cmu.edu/homepage/computing/2013/summer/apportable-entrepreneurship.shtml), 2013.
31. Shumeet Baluja and Ian Fischer, [Adversarial Transformation Networks: Learning to Generate Adversarial Examples](https://arxiv.org/abs/1703.09387), arXiv:1703.09387, March 2017.
32. Kuang-Huei Lee, Ian Fischer, Yueh-Hua Wu, Dave Marwood, Shumeet Baluja, Dale Schuurmans and Xinyun Chen, [Evolving Deeper LLM Thinking](https://arxiv.org/abs/2501.09891), arXiv:2501.09891, January 2025.
33. Alexander A. Alemi, Ian Fischer, Joshua V. Dillon and Kevin Murphy, [Deep Variational Information Bottleneck](https://arxiv.org/abs/1612.00410), arXiv:1612.00410, December 2016.
34. Ian Fischer, [The Conditional Entropy Bottleneck](https://arxiv.org/abs/2002.05379), arXiv:2002.05379, February 2020.
35. Danijar Hafner, Timothy Lillicrap, Ian Fischer, Ruben Villegas, David Ha, Honglak Lee and James Davidson, [Learning Latent Dynamics for Planning from Pixels](https://arxiv.org/abs/1811.04551), arXiv:1811.04551, November 2018.
36. Kuang-Huei Lee, Xinyun Chen, Hiroki Furuta, John Canny and Ian Fischer, [A Human-Inspired Reading Agent with Gist Memory of Very Long Contexts](https://arxiv.org/abs/2402.09727), arXiv:2402.09727, February 2024.
37. Shumeet Baluja, [The Silicon Jungle (book announcement)](https://mailman.srv.cs.cmu.edu/pipermail/connectionists/2011-April/024490.html), Connectionists mailing list, April 28, 2011.
38. Sahar Vahdati, Andrei Aioanei, Haridhra Suresh and Jens Lehmann, [The ARC of Progress towards AGI: A Living Survey of Abstraction and Reasoning](https://arxiv.org/abs/2603.13372), arXiv:2603.13372, March 2026.
39. ARC Prize, [ARC-AGI Leaderboard](https://arcprize.org/leaderboard), accessed October 27, 2026.
40. Debarghya Das, [The Ultimate Neolab List](https://debarghyadas.com/neolabs/), accessed October 27, 2026.