Guest User

Untitled

a guest
Aug 11th, 2026
62
0
Never
Not a member of Pastebin yet? Sign Up, it unlocks many cool features!
text 14.30 KB | None | 0 0
  1. # SOVEREIGN FILIPINO & ON-DEVICE AI — 2026 STATE OF PLAY AND THE SOLO-BUILDER OPPORTUNITY
  2.  
  3. **Prepared:** 2026-08-11 · **Method:** HoardCore DeepResearch (Hardcore Research Loop; 6 dedicated discovery/recall passes + reuse of 15 Pax-Silica passes) · **Lens:** solo Filipino NLP/edge-AI builder · **Vault:** `hoardcore_data/vault.db`
  4.  
  5. > **Provenance:** `[V]` = confirmed in full primary text stored in the HoardCore vault (grounding_phai_1..2, grounding_pax_8/11/12). `[E]` = captured this session but not vault-ingested. `[H]` = author reasoning. Every number tagged; unverifiable claims flagged.
  6.  
  7. ---
  8.  
  9. ## 1. EXECUTIVE SUMMARY
  10.  
  11. - **The frontier models are bad at Philippine languages — and that is a durable, currently-unpriced market gap, not a temporary one.** FilBench, the first systematic Filipino/Tagalog/Cebuano LLM benchmark, evaluates 20+ models across 4 categories / 12 subtasks and finds **text generation is the consistent weak spot across all models** — instruction misalignment, verbosity, and hallucinating a different language `[V#2][V#5]`. Meanwhile open-weight SEA models (SEA-LION, SeaLLM) outperform on Filipino tasks and are cheaper to run than commercial baselines, with fine-tuning on SEA data yielding only **+2–3% on FilBench** — meaning the low-hanging fruit (better data, better eval, on-device adaptation) is wide open `[V#2]`.
  12. - **On-device inference is now technically solved enough to build on — quantization to <10 MB with ~1–2% accuracy loss, running on a 4 GB Raspberry Pi or midrange Android** `[V#6][V#8]`. The live proof is **SugbuVoice**, an offline Cebuano voicebot on a Raspberry Pi handling LGU service requests, with **three pilots signed in Q1 2026** and **Hiligaynon slated mid-2026** `[V#7]`. That is the same product category, architecture (offline, on-device, low-resource language), and buyer (public sector) as your FiliSenti/HiliSenti work.
  13. - **Your single highest-leverage move:** build and publish the missing link between these two facts — a **benchmark-validated, INT8-quantized, on-device Hiligaynon/Tagalog NLP + small-LM voicebot** — funded through the **DOST-PCIEERD research-grant / DTI MSME-grant (₱100K–₱10M) lane**, with government wind at your back (KWF is officially embracing AI for indigenous languages `[V#10]`). The public/edge lane is funded and undersupplied today; Pax Silica's "trusted edge software stack" is a 2028–2030 optionality on top `[H]`.
  14.  
  15. ---
  16.  
  17. ## 2. THE LANDSCAPE — WHAT EXISTS, WHAT'S MISSING (2026)
  18.  
  19. ### 2.1 What exists (foundations, all open/curated)
  20. - **DOST-ASTI `iTANONG-DS`** on Hugging Face: BERT & RoBERTa for Cebuano + Tagalog, GPT-2 Tagalog, BigBird Tagalog, and a **RoBERTa Tagalog sentiment model**; the broader `iTANONG` project targets Filipino/Taglish database querying for public information access `[V#12]`.
  21. - **Established open base models:** `jcblaise/roberta-tagalog-base` (TLUnified corpus) and `jfernandez/cebuano-filipino-language-model` `[V]`.
  22. - **Toolkits:** **CalamanCy** (open-source Tagalog SpaCy toolkit: POS, dependency parsing, NER) and **FiLLM** (optimizing a regional open model for Filipino tasks) `[V#12]`.
  23. - **Datasets/benchmarks:** **FilBench** (Tagalog/Filipino/Cebuano, 4 categories / 12 subtasks, MCF + ROUGE-L) `[V#1]`; curated resources like INCLUDE, KALAHI, Dengue Filipino, BalitaNLP, CebuaNER, Cebuano Readability Corpus, Belebele `[V#5]`.
  24. - **Sovereign-LLM momentum:** **Budots Media** is building a sovereign Cebuano/Tagalog **multimodal** model on a decade of rights-cleared media, arguing 180+ PH languages are near-absent from frontier models (Hiligaynon/Waray/Bikol "barely at all") `[V#13]`; Iloilo-based **Cedrick Alegroso** shipped a **Hiligaynon** model trained on ~80k tokens of Bombo Radyo text — proof that resource-constrained builders can do this `[V#12]`.
  25. - **Policy tailwind:** **KWF** (Komisyon sa Wikang Filipino) formally embracing AI to strengthen Filipino/indigenous languages (ABS-CBN, Jul 15 2026) `[V#10]`.
  26.  
  27. ### 2.2 What's missing (the gap = your market)
  28. - **Generation quality** — the category most needed for voicebots/chatbots/assistants — is the *weakest* across all 20+ models `[V#2][V#5]`.
  29. - **On-device, quantized, productized Filipino models** tuned for a specific use-case (LGU/coop/public service) are scarce; most open work is research-grade, not deployable `[H]`.
  30. - **Hiligaynon, and the Visayan mother tongues beyond Cebuano, are under-served** relative to Tagalog `[V#13][V#12]` — yet are the languages people actually speak in Negros, Iloilo, Bacolod `[H]`.
  31. - **Evaluation for deployment** (latency, memory, INT8 accuracy on-device for Filipino tasks) does not exist as a public artifact `[H]`.
  32.  
  33. ---
  34.  
  35. ## 3. THE EVALUATION REALITY (WHY YOUR NICHE IS UNPRICED)
  36.  
  37. - FilBench measures **Cultural Knowledge, Classical NLP, Reading Comprehension, and Generation** across **Filipino, Tagalog, Cebuano**; 20+ SOTA LLMs; uses MCF + ROUGE-L `[V#1][V#5]`.
  38. - **SEA-specific open-weight LLMs** (SEA-LION, SeaLLM) achieve the highest FilBench scores among their size class; **GPT-4o** remains a strong closed-source baseline; **open-weight models are cheaper to run** `[V#2]`.
  39. - Fine-tuning on SEA-specific data yields **only +2–3% FilBench gains** — i.e., current fine-tuning is not yet capturing the language-specific structure; this is a *technique + data* frontier, not a compute-arms-race frontier `[V#2]`.
  40. - Filipino has **>28M speakers in the PH and >2M abroad** (PSA 2020) — a large, underserved, commercially relevant language market `[V#1]`.
  41. - **Reading/generation deficits** are named as the explicit guide for "future region-specific model improvements" `[V#3]` — a direct research invitation.
  42.  
  43. ---
  44.  
  45. ## 4. THE ON-DEVICE / EDGE OPPORTUNITY (YOUR STACK, VALIDATED)
  46.  
  47. - Quantized edge models fit in **<10 MB** and run at **~40–80 ms/frame on a Raspberry Pi 4/5**, **<30 ms on a modern smartphone**, using **<64 MB peak RAM**, with **~1–2% accuracy vs float baseline** after quantization `[V#8]`.
  48. - Small language models + edge AI is a named **2026 shift to "local intelligence"** `[V#9]`; quantization strategies for local SLM inference are being systematized `[V#6]`.
  49. - **SugbuVoice** = production proof: a fine-tuned small LM on a **4 GB Raspberry Pi at a barangay hall**, handling routine LGU requests in **Cebuano**; **three pilots Q1 2026**, **Hiligaynon planned mid-2026** `[V#7]`.
  50. - Implication: your **ONNX + INT8 + Flutter + offline-first** stack maps directly onto this; the missing piece is a *Filipino/Hiligaynon* instance + a benchmark/eval artifact `[H]`.
  51.  
  52. ---
  53.  
  54. ## 5. THE FUNDING LANE (NON-DILUTIVE, APPLY THIS QUARTER)
  55.  
  56. - **DOST-PCIEERD research grants** (R&D) — directly relevant to a solo ML builder; can be layered with BOI status and DTI subsidies (mind anti-double-dipping clauses) `[V#11]`.
  57. - **DTI startup/MSME grants: ₱100K–₱10M non-dilutive**, weighted toward innovation & digitalization; DTI prioritizes **food security + digital-transformation** ventures, with **Clark, Cebu, Davao innovation zones** expedited `[V#11]`.
  58. - Priority sectors named: **agritech, fintech, healthtech, BPO**; **climate-adaptation & RE-microgrid proposals get priority weighting** `[V#11]` — an "AI for energy / public service" framing slots in.
  59. - PEZA **4–8 yr income-tax holidays** (export-oriented) and BOI strategic-industry exemptions exist but suit manufacturing/export more than a solo product `[H]`.
  60. - PH AI-startup funding reality: **small and focused** — most deals in **Filipino-language NLP, AI-augmented BPO, fintech underwriting**; funding restarted late-2025/2026 after a 2023–24 survival mode, **not at 2021 valuations** `[V#14]`. Consumer AI lags B2B — so public-sector/B2B edge deployments are the funded lane `[H]`.
  61.  
  62. ---
  63.  
  64. ## 6. THE PAX SILICA OPTIONALITY (2028–2030, DO NOT BET THE PLAN ON IT)
  65.  
  66. - **Edge AI Package** ($200M, US State Dept, Feb 2026): subsidizes low-cost "AI-ready" Android/iOS handsets in the Indo-Pacific through MNOs/OEMs; **90-day window closed ~May 2026**; **no public awardees as of Aug 2026**; explicitly aims for a **"trusted AI software stack"** and developer/AI-innovation tooling `[V#15][E]`.
  67. - The package's own money is a **hardware-distribution subsidy, not a model-builder grant** — but the "AI-ready trusted handsets" it seeds create a downstream market for **preloaded, offline-first, localized AI apps** `[H]`.
  68. - **ESZ timeline:** framework agreement target **~Nov 2026**; detailed planning **2027**; construction **~2028**; first tenant (reported Foxconn, none definitive) due-diligence ongoing; jobs figure (130k–190k) **not disaggregated into software** `[V#16][V#17]`.
  69. - **Net:** treat Pax Silica as optionality that *amplifies* the sovereign-edge story (trusted devices running trusted localized AI), not as a near-term revenue path `[H]`.
  70.  
  71. ---
  72.  
  73. ## 7. STRATEGIC RECOMMENDATION FOR A SOLO BUILDER
  74.  
  75. **Build the missing artifact, then sell it in the funded public/edge lane.** Concretely:
  76.  
  77. 1. **Ship a benchmark-validated, INT8 on-device Filipino/Hiligaynon NLP + small-LM voicebot** (mirror SugbuVoice's use-case; target Hiligaynon/Tagalog where coverage is thinnest). Publish it with a **FilBench-style eval** and **on-device latency/memory numbers** — this is the artifact the field lacks `[H]`.
  78. 2. **Apply to DOST-PCIEERD and DTI MSME (₱100K–₱10M) grants** framing it as "on-device, privacy-preserving Filipino NLP for public-service delivery" — matches priority sectors and the sovereignty thesis `[V#11]`.
  79. 3. **Monetize in the proven existing market:** LGU/coop/BPO B2B edge deployments (the SugbuVoice proof `[V#7]`), while keeping a Pax-Silica "trusted edge software stack" plug-in package ready `[H]`.
  80.  
  81. **The bet:** independent sovereign Filipino edge-AI, built to be Pax-Silica-ready but not Pax-Silica-dependent `[H]`.
  82.  
  83. ---
  84.  
  85. ## 8. TIMELINE & TRIGGERS
  86.  
  87. | When | Signal | Source |
  88. |---|---|---|
  89. | Now (Q3 2026) | DOST-PCIEERD / DTI grant windows open; KWF AI-for-languages active | `[V#10][V#11]` |
  90. | ~Mid 2026 | SugbuVoice Hiligaynon extension (watch as market validation) | `[V#7]` |
  91. | ~Nov 2026 | Pax Silica framework signing (optionality trigger) | `[V#17]` |
  92. | 2027 | ESZ detailed planning | `[V#18]` |
  93. | ~2028 | ESZ construction / first-tenant deals | `[V#17]` |
  94. | 2028–2030 | First "trusted edge software stack" opportunities (if deals hold) | `[H]` |
  95.  
  96. **What to watch:** new FilBench submissions; DOST-ASTI iTANONG releases; KWF initiatives; SugbuVoice's Hiligaynon rollout; any Edge AI Package awardees; DICT Digital Cities / CAIR (NAISR 2.0) grant calls `[V]`.
  97.  
  98. ---
  99.  
  100. ## SOURCE LINKS / CITATIONS
  101.  
  102. Numbered tags `[V#N]`/`[E#N]` resolve below. `[V]` = stored in the HoardCore vault; `[E]` = captured this session, not ingested.
  103.  
  104. [#1] arXiv 2508.03523 — FilBench: a Filipino NLP benchmark for LLMs (Tagalog, Filipino, Cebuano) — https://arxiv.org/pdf/2508.03523
  105. [#2] Hugging Face Blog via BrainAI — FilBench: Can LLMs understand and generate Filipino? — https://brainai.pro/news/en/2025/08/12/filbench-can-llms-understand-and-generate-filipino-a-deep-dive-into-tagalog-and-/
  106. [#3] EmergentMind — FilBench: Filipino NLP Benchmark for LLMs (structure & findings) — https://www.emergentmind.com/topics/filbench
  107. [#4] BrainAI — FilBench: Filipino Language Evaluation Suite for LLMs — https://brainai.pro/resources/en/2025/08/12/filbench-filipino-language-evaluation-suite-for-llms-tagalog-filipino-cebuano/
  108. [#5] FilBench paper (task categories/datasets: INCLUDE, KALAHI, Dengue Filipino, BalitaNLP, CebuaNER, Cebuano Readability, Belebele) — via [#1]
  109. [#6] martinuke0.github.io — Optimizing small language models for local edge inference: quantization 2026 — https://martinuke0.github.io/posts/2026-03-26-optimizing-small-language-models-for-local-edge-inference-a-guide-to-quantization-in-2026/
  110. [#7] ImpactHub PH — SugbuVoice: offline Cebuano voicebot on Raspberry Pi (3 LGU pilots Q1 2026; Hiligaynon mid-2026) — https://impacthub.ph/article-edge-ai-philippines.html
  111. [#8] ASOasis — Edge AI on-device inference tutorial (TFLite; <10 MB; <30 ms phone; ~1–2% drop) — https://asoasis.tech/articles/2026-05-02-0853-edge-ai-on-device-inference-tutorial/
  112. [#9] Zylos.ai — Small Language Models and Edge AI: the 2026 shift to local intelligence — https://zylos.ai/research/2026-02-07-small-language-models-edge-ai/
  113. [#10] ABS-CBN — KWF embraces AI to strengthen Filipino, indigenous languages — https://www.abs-cbn.com/news/technology/2026/7/15/kwf-embraces-ai-to-strengthen-filipino-indigenous-languages-1707
  114. [#11] Pertama Partners — Philippines Funding Guide 2026 (DTI ₱100K–₱10M grants; DOST-PCIEERD; priority sectors) — https://www.pertamapartners.com/funding/philippines-funding-guide-2026
  115. [#12] Nordis — Filipino Developers Are Building Local AI Tools (DOST-ASTI iTANONG; CalamanCy; FiLLM; Hiligaynon model) — https://nordis.net/2026/05/15/article/opinion/filipino-developers-are-building-local-ai-tools/
  116. [#13] Budots Media — Sovereign Filipino LLM (Cebuano/Tagalog multimodal; 180+ languages; Hiligaynon/Waray/Bikol absent) — https://barts83eu.github.io/budotsmediaph-202606-fable5/stories/ai-filipino-llm/
  117. [#14] WebGeek — Filipino AI startups funding 2026 (scene small; NLP/BPO/fintech; funding restarted) — https://webgeek.ph/news/filipino-ai-startups-funding-2026/
  118. [#15] CNBC TV18 — US launches $200M Edge AI Package (Pax Silica; trusted AI software stack) — https://www.cnbctv18.com/technology/us-launches-usd-200-million-edge-ai-package-to-boost-secure-smartphones-in-indo-pacific-region-ws-el-19854293.htm
  119. [#16] Hugging Face — jcblaise/roberta-tagalog-base — https://huggingface.co/jcblaise/roberta-tagalog-base
  120. [#17] Philstar — BCDA: 30+ firms; November framework; 130k–190k jobs; construction 2028 — https://www.philstar.com/business/2026/07/24/2544182/bcda-0ver-30-firms-interested-pax-silica-project
  121. [#18] Ibrixon — Pax Silica & New Clark City plan (2027 planning; 2028 construction) — https://www.ibrixon.com/post/pax-silica-and-new-clark-city-philippines-plan-for-an-ai-and-semiconductor-hub
  122.  
  123. **Flagged:** Edge AI Package awardees — none public as of 2026-08-11 (`[E]`). Job-count disaggregation — unpublished (`[H]`). "Generation is weakest" is `[V]` (FilBench), but *market impact* framing is `[H]`. Null results = "no public record located this session."
  124.  
  125. *Generated by HoardCore DeepResearch. Grounding: artifacts/2026-08-11/grounding_phai_1..2.md + grounding_pax_8/11/12.md. Vault: hoardcore_data/vault.db.*
  126.  
Add Comment
Please, Sign In to add comment